Merge branch 'development' into triangular_solvers

30 files changed, 67 insertions, 110 deletions

diff --git a/src/routines/common.hpp b/src/routines/common.hpp
index 8046c0be..bdea0086 100644
--- a/src/routines/common.hpp
+++ b/src/routines/common.hpp
@@ -19,8 +19,8 @@
 #include <string>
 #include <vector>
 
-#include "clblast.h"
 #include "clpp11.hpp"
+#include "clblast.h"
 #include "database/database.hpp"
 
 namespace clblast {
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) {
diff --git a/src/routines/level2/xgemv.cpp b/src/routines/level2/xgemv.cpp
index 52e66de6..7d2e5f60 100644
--- a/src/routines/level2/xgemv.cpp
+++ b/src/routines/level2/xgemv.cpp
@@ -123,8 +123,7 @@ void Xgemv<T>::MatVec(const Layout layout, const Transpose a_transpose,
   }
 
   // Retrieves the Xgemv 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
   kernel.SetArgument(0, static_cast<int>(m_real));
diff --git a/src/routines/level2/xger.cpp b/src/routines/level2/xger.cpp
index d16ebd11..9ec156a1 100644
--- a/src/routines/level2/xger.cpp
+++ b/src/routines/level2/xger.cpp
@@ -53,8 +53,7 @@ void Xger<T>::DoGer(const Layout layout,
   TestVectorY(n, y_buffer, y_offset, y_inc);
 
   // Retrieves the kernel from the compiled binary
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, "Xger");
+  auto kernel = Kernel(program_, "Xger");
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(a_one));
diff --git a/src/routines/level2/xher.cpp b/src/routines/level2/xher.cpp
index 6c334e63..ba12a3ef 100644
--- a/src/routines/level2/xher.cpp
+++ b/src/routines/level2/xher.cpp
@@ -67,8 +67,7 @@ void Xher<T,U>::DoHer(const Layout layout, const Triangle triangle,
   const auto matching_alpha = GetAlpha(alpha);
 
   // Retrieves the kernel from the compiled binary
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, "Xher");
+  auto kernel = Kernel(program_, "Xher");
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(n));
diff --git a/src/routines/level2/xher2.cpp b/src/routines/level2/xher2.cpp
index 11e2c871..a420e693 100644
--- a/src/routines/level2/xher2.cpp
+++ b/src/routines/level2/xher2.cpp
@@ -54,8 +54,7 @@ void Xher2<T>::DoHer2(const Layout layout, const Triangle triangle,
   TestVectorY(n, y_buffer, y_offset, y_inc);
 
   // Retrieves the kernel from the compiled binary
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, "Xher2");
+  auto kernel = Kernel(program_, "Xher2");
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(n));
diff --git a/src/routines/level2/xtrsv.cpp b/src/routines/level2/xtrsv.cpp
index b0e4c5ae..d5d009ff 100644
--- a/src/routines/level2/xtrsv.cpp
+++ b/src/routines/level2/xtrsv.cpp
@@ -37,9 +37,6 @@ void Xtrsv<T>::Substitution(const Layout layout, const Triangle triangle,
 
   if (n > db_["TRSV_BLOCK_SIZE"]) { throw BLASError(StatusCode::kUnexpectedError); };
 
-  // Retrieves the program from the cache
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), "TRSV");
-
   // Translates CLBlast arguments to 0/1 integers for the OpenCL kernel
   const auto is_unit_diagonal = (diagonal == Diagonal::kNonUnit) ? 0 : 1;
   const auto is_transposed = ((a_transpose == Transpose::kNo && layout == Layout::kColMajor) ||
@@ -52,7 +49,7 @@ void Xtrsv<T>::Substitution(const Layout layout, const Triangle triangle,
 
   // Retrieves the kernel from the compiled binary
   const auto kernel_name = (is_upper) ? "trsv_backward" : "trsv_forward";
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(n));
@@ -94,9 +91,6 @@ void Xtrsv<T>::DoTrsv(const Layout layout, const Triangle triangle,
   TestMatrixA(n, n, a_buffer, a_offset, a_ld);
   TestVectorX(n, b_buffer, b_offset, b_inc);
 
-  // Retrieves the program from the cache
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), "TRSV");
-
   // Creates a copy of B to avoid overwriting input while computing output
   // TODO: Make x with 0 offset and unit increment by creating custom copy-to and copy-from kernels
   const auto x_offset = b_offset;
@@ -108,7 +102,7 @@ void Xtrsv<T>::DoTrsv(const Layout layout, const Triangle triangle,
   // Fills the output buffer with zeros
   auto eventWaitList = std::vector<Event>();
   auto fill_vector_event = Event();
-  FillVector(queue_, device_, program, db_, fill_vector_event.pointer(), eventWaitList,
+  FillVector(queue_, device_, program_, db_, fill_vector_event.pointer(), eventWaitList,
              n, x_inc, x_offset, x_buffer, ConstantZero<T>());
   fill_vector_event.WaitForCompletion();
 
diff --git a/src/routines/level2/xtrsv.hpp b/src/routines/level2/xtrsv.hpp
index dc3f32f0..67e626a1 100644
--- a/src/routines/level2/xtrsv.hpp
+++ b/src/routines/level2/xtrsv.hpp
@@ -27,11 +27,11 @@ class Xtrsv: public Xgemv<T> {
  public:
 
   // Uses the generic matrix-vector routine
-  using Xgemv<T>::routine_name_;
   using Xgemv<T>::queue_;
   using Xgemv<T>::context_;
   using Xgemv<T>::device_;
   using Xgemv<T>::db_;
+  using Xgemv<T>::program_;
   using Xgemv<T>::DoGemv;
 
   // Constructor
diff --git a/src/routines/level3/xgemm.cpp b/src/routines/level3/xgemm.cpp
index 0015b629..7bd388c1 100644
--- a/src/routines/level3/xgemm.cpp
+++ b/src/routines/level3/xgemm.cpp
@@ -150,9 +150,6 @@ void Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k,
   const auto c_one_i = (c_want_rotated) ? n_ceiled : m_ceiled;
   const auto c_two_i = (c_want_rotated) ? m_ceiled : n_ceiled;
 
-  // 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 = a_one == a_one_i && a_two == a_two_i && a_ld == a_one && a_offset == 0 &&
                    a_do_transpose == false && a_conjugate == false;
@@ -178,7 +175,7 @@ void Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k,
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
                            a_one, a_two, a_ld, a_offset, a_buffer,
                            a_one_i, a_two_i, a_one_i, 0, a_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, a_do_transpose, a_conjugate);
     eventWaitList.push_back(eventProcessA);
   }
@@ -189,7 +186,7 @@ void Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k,
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList,
                            b_one, b_two, b_ld, b_offset, b_buffer,
                            b_one_i, b_two_i, b_one_i, 0, b_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, b_do_transpose, b_conjugate);
     eventWaitList.push_back(eventProcessB);
   }
@@ -200,13 +197,13 @@ void Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k,
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
                            c_one, c_two, c_ld, c_offset, c_buffer,
                            c_one_i, c_two_i, c_one_i, 0, c_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, c_do_transpose, false);
     eventWaitList.push_back(eventProcessC);
   }
 
   // Retrieves the Xgemm kernel from the compiled binary
-  auto kernel = Kernel(program, "Xgemm");
+  auto kernel = Kernel(program_, "Xgemm");
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(m_ceiled));
@@ -236,7 +233,7 @@ void Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k,
     PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList,
                            c_one_i, c_two_i, c_one_i, 0, c_temp,
                            c_one, c_two, c_ld, c_offset, c_buffer,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            false, c_do_transpose, false);
   }
 }
@@ -255,13 +252,10 @@ void Xgemm<T>::GemmDirect(const size_t m, const size_t n, const size_t k,
                           const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose,
                           const bool a_conjugate, const bool b_conjugate) {
 
-  // Loads the program from the database
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-
   // Retrieves the proper XgemmDirect kernel from the compiled binary
   const auto name = (a_do_transpose) ? (b_do_transpose ? "XgemmDirectTT" : "XgemmDirectTN") :
                                        (b_do_transpose ? "XgemmDirectNT" : "XgemmDirectNN");
-  auto kernel = Kernel(program, name);
+  auto kernel = Kernel(program_, name);
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(m));
diff --git a/src/routines/level3/xhemm.cpp b/src/routines/level3/xhemm.cpp
index e5b1502a..8629f3de 100644
--- a/src/routines/level3/xhemm.cpp
+++ b/src/routines/level3/xhemm.cpp
@@ -58,8 +58,7 @@ void Xhemm<T>::DoHemm(const Layout layout, const Side side, const Triangle trian
 
   // Creates a general matrix from the hermitian matrix to be able to run the regular Xgemm
   // routine afterwards
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the arguments for the hermitian-to-squared kernel
   kernel.SetArgument(0, static_cast<int>(k));
diff --git a/src/routines/level3/xhemm.hpp b/src/routines/level3/xhemm.hpp
index 2385706e..7c011915 100644
--- a/src/routines/level3/xhemm.hpp
+++ b/src/routines/level3/xhemm.hpp
@@ -30,6 +30,7 @@ class Xhemm: public Xgemm<T> {
   using Xgemm<T>::queue_;
   using Xgemm<T>::context_;
   using Xgemm<T>::device_;
+  using Xgemm<T>::program_;
   using Xgemm<T>::db_;
   using Xgemm<T>::DoGemm;
 
diff --git a/src/routines/level3/xher2k.cpp b/src/routines/level3/xher2k.cpp
index ee3bb8b8..2aed2781 100644
--- a/src/routines/level3/xher2k.cpp
+++ b/src/routines/level3/xher2k.cpp
@@ -81,9 +81,6 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
   // 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 a1_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 &&
                     ab_rotated == false && ab_conjugate == false;
@@ -116,7 +113,7 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA1.pointer(), emptyEventList,
                            ab_one, ab_two, a_ld, a_offset, a_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, a1_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, ab_rotated, ab_conjugate);
     eventWaitList.push_back(eventProcessA1);
   }
@@ -125,7 +122,7 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA2.pointer(), emptyEventList,
                            ab_one, ab_two, a_ld, a_offset, a_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, a2_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, ab_rotated, !ab_conjugate);
     eventWaitList.push_back(eventProcessA2);
   }
@@ -134,7 +131,7 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB1.pointer(), emptyEventList,
                            ab_one, ab_two, b_ld, b_offset, b_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, b1_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, ab_rotated, ab_conjugate);
     eventWaitList.push_back(eventProcessB1);
   }
@@ -143,7 +140,7 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB2.pointer(), emptyEventList,
                            ab_one, ab_two, b_ld, b_offset, b_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, b2_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, ab_rotated, !ab_conjugate);
     eventWaitList.push_back(eventProcessB2);
   }
@@ -154,12 +151,12 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
   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));
@@ -201,7 +198,7 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
   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, true);
 }
 
diff --git a/src/routines/level3/xherk.cpp b/src/routines/level3/xherk.cpp
index ae8e9324..d982859e 100644
--- a/src/routines/level3/xherk.cpp
+++ b/src/routines/level3/xherk.cpp
@@ -79,9 +79,6 @@ void Xherk<T,U>::DoHerk(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 = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 &&
                    a_rotated == false && a_conjugate == false;
@@ -109,7 +106,7 @@ void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Tran
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
                            a_one, a_two, a_ld, a_offset, a_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, a_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, a_rotated, a_conjugate);
     eventWaitList.push_back(eventProcessA);
   }
@@ -118,7 +115,7 @@ void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Tran
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList,
                            a_one, a_two, a_ld, a_offset, a_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, b_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, a_rotated, b_conjugate);
     eventWaitList.push_back(eventProcessB);
   }
@@ -129,12 +126,12 @@ void Xherk<T,U>::DoHerk(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));
@@ -163,7 +160,7 @@ void Xherk<T,U>::DoHerk(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, true);
 }
 
diff --git a/src/routines/level3/xsymm.cpp b/src/routines/level3/xsymm.cpp
index d7f771d1..969edfc8 100644
--- a/src/routines/level3/xsymm.cpp
+++ b/src/routines/level3/xsymm.cpp
@@ -30,12 +30,12 @@ Xsymm<T>::Xsymm(Queue &queue, EventPointer event, const std::string &name):
 // The main routine
 template <typename T>
 void Xsymm<T>::DoSymm(const Layout layout, const Side side, const Triangle triangle,
-                            const size_t m, const size_t n,
-                            const T alpha,
-                            const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
-                            const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld,
-                            const T beta,
-                            const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld) {
+                      const size_t m, const size_t n,
+                      const T alpha,
+                      const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
+                      const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld,
+                      const T beta,
+                      const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld) {
 
   // Makes sure all dimensions are larger than zero
   if ((m == 0) || (n == 0) ) { throw BLASError(StatusCode::kInvalidDimension); }
@@ -58,8 +58,7 @@ void Xsymm<T>::DoSymm(const Layout layout, const Side side, const Triangle trian
 
   // Creates a general matrix from the symmetric matrix to be able to run the regular Xgemm
   // routine afterwards
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the arguments for the symmetric-to-squared kernel
   kernel.SetArgument(0, static_cast<int>(k));
diff --git a/src/routines/level3/xsymm.hpp b/src/routines/level3/xsymm.hpp
index ee965364..7a584560 100644
--- a/src/routines/level3/xsymm.hpp
+++ b/src/routines/level3/xsymm.hpp
@@ -32,6 +32,7 @@ class Xsymm: public Xgemm<T> {
   using Xgemm<T>::queue_;
   using Xgemm<T>::context_;
   using Xgemm<T>::device_;
+  using Xgemm<T>::program_;
   using Xgemm<T>::db_;
   using Xgemm<T>::DoGemm;
 
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);
 }
 
diff --git a/src/routines/level3/xsyrk.cpp b/src/routines/level3/xsyrk.cpp
index bd6c4b25..9588c28c 100644
--- a/src/routines/level3/xsyrk.cpp
+++ b/src/routines/level3/xsyrk.cpp
@@ -74,9 +74,6 @@ void Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const Transp
   // 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 = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 &&
                    a_rotated == false;
@@ -97,7 +94,7 @@ void Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const Transp
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
                            a_one, a_two, a_ld, a_offset, a_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, a_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, a_rotated, false);
     eventWaitList.push_back(eventProcessA);
   }
@@ -108,12 +105,12 @@ void Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const Transp
   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));
@@ -142,7 +139,7 @@ void Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const Transp
   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);
 }
 
diff --git a/src/routines/level3/xtrmm.cpp b/src/routines/level3/xtrmm.cpp
index ed810e72..02c295ac 100644
--- a/src/routines/level3/xtrmm.cpp
+++ b/src/routines/level3/xtrmm.cpp
@@ -70,8 +70,7 @@ void Xtrmm<T>::DoTrmm(const Layout layout, const Side side, const Triangle trian
 
   // Creates a general matrix from the triangular matrix to be able to run the regular Xgemm
   // routine afterwards
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the arguments for the triangular-to-squared kernel
   kernel.SetArgument(0, static_cast<int>(k));
diff --git a/src/routines/level3/xtrmm.hpp b/src/routines/level3/xtrmm.hpp
index 967bf132..e77b7214 100644
--- a/src/routines/level3/xtrmm.hpp
+++ b/src/routines/level3/xtrmm.hpp
@@ -31,6 +31,7 @@ class Xtrmm: public Xgemm<T> {
   using Xgemm<T>::queue_;
   using Xgemm<T>::context_;
   using Xgemm<T>::device_;
+  using Xgemm<T>::program_;
   using Xgemm<T>::db_;
   using Xgemm<T>::DoGemm;
 
diff --git a/src/routines/level3/xtrsm.cpp b/src/routines/level3/xtrsm.cpp
index 8061b508..3a910261 100644
--- a/src/routines/level3/xtrsm.cpp
+++ b/src/routines/level3/xtrsm.cpp
@@ -79,9 +79,8 @@ void Xtrsm<T>::DoTrsm(const Layout layout, const Side side, const Triangle trian
 
   // Fills the output buffer with zeros
   auto eventWaitList = std::vector<Event>();
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), "TRSM");
   auto fill_matrix_event = Event();
-  FillMatrix(queue_, device_, program, db_, fill_matrix_event.pointer(), eventWaitList,
+  FillMatrix(queue_, device_, program_, db_, fill_matrix_event.pointer(), eventWaitList,
              x_one, x_ld, x_offset, x_buffer, ConstantZero<T>());
   fill_matrix_event.WaitForCompletion();
 
diff --git a/src/routines/level3/xtrsm.hpp b/src/routines/level3/xtrsm.hpp
index 288e9d11..b9d5432a 100644
--- a/src/routines/level3/xtrsm.hpp
+++ b/src/routines/level3/xtrsm.hpp
@@ -26,11 +26,11 @@ class Xtrsm: public Xgemm<T> {
  public:
 
   // Uses methods and variables the Xgemm routine
-  using Xgemm<T>::routine_name_;
   using Xgemm<T>::queue_;
   using Xgemm<T>::context_;
   using Xgemm<T>::device_;
   using Xgemm<T>::db_;
+  using Xgemm<T>::program_;
   using Xgemm<T>::DoGemm;
 
   // Constructor
diff --git a/src/routines/levelx/xinvert.cpp b/src/routines/levelx/xinvert.cpp
index ffee9b7c..696e694a 100644
--- a/src/routines/levelx/xinvert.cpp
+++ b/src/routines/levelx/xinvert.cpp
@@ -69,18 +69,15 @@ void Xinvert<T>::InvertMatrixDiagonalBlocks(const Layout layout, const Triangle
                          (triangle == Triangle::kLower && layout == Layout::kRowMajor));
   const auto name_postfix = (is_upper) ? "Upper" : "Lower";
 
-  // Retrieves the program from the cache
-  auto event_wait_list = std::vector<Event>();
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), "INVERT");
-
   // Fills the output buffer with zeros
+  auto event_wait_list = std::vector<Event>();
   auto fill_matrix_event = Event();
-  FillMatrix(queue_, device_, program, db_, fill_matrix_event.pointer(), event_wait_list,
+  FillMatrix(queue_, device_, program_, db_, fill_matrix_event.pointer(), event_wait_list,
              num_blocks * block_size, block_size, 0, dest, ConstantZero<T>());
   event_wait_list.push_back(fill_matrix_event);
 
   // Inverts the diagonal IB by IB inner blocks of the matrix: one block per work-group
-  auto kernel = Kernel(program, "InvertDiagonalBlock");
+  auto kernel = Kernel(program_, "InvertDiagonalBlock");
   kernel.SetArgument(0, static_cast<int>(n));
   kernel.SetArgument(1, src());
   kernel.SetArgument(2, static_cast<int>(offset));
@@ -110,7 +107,7 @@ void Xinvert<T>::InvertMatrixDiagonalBlocks(const Layout layout, const Triangle
     const auto global = std::vector<size_t>{(current_size/local[1]), npages*(current_size/16)*local[1]};
 
     // Part 1
-    auto kernel1 = Kernel(program, "TripleMatMul" + ToString(current_size) + "Part1" + name_postfix);
+    auto kernel1 = Kernel(program_, "TripleMatMul" + ToString(current_size) + "Part1" + name_postfix);
     kernel1.SetArgument(0, static_cast<int>(n));
     kernel1.SetArgument(1, src());
     kernel1.SetArgument(2, static_cast<int>(offset));
@@ -125,7 +122,7 @@ void Xinvert<T>::InvertMatrixDiagonalBlocks(const Layout layout, const Triangle
 
     // Part 2
     const bool is_last_kernel = (current_size * 2 >= block_size);
-    auto kernel2 = Kernel(program, "TripleMatMul" + ToString(current_size) + "Part2" + name_postfix);
+    auto kernel2 = Kernel(program_, "TripleMatMul" + ToString(current_size) + "Part2" + name_postfix);
     kernel2.SetArgument(0, static_cast<int>(n));
     kernel2.SetArgument(1, dest());
     kernel2.SetArgument(2, static_cast<int>(current_size));
diff --git a/src/routines/levelx/xomatcopy.cpp b/src/routines/levelx/xomatcopy.cpp
index 875ca7d2..4ae8c056 100644
--- a/src/routines/levelx/xomatcopy.cpp
+++ b/src/routines/levelx/xomatcopy.cpp
@@ -65,14 +65,11 @@ void Xomatcopy<T>::DoOmatcopy(const Layout layout, const Transpose a_transpose,
   TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld);
   TestMatrixB(b_one, b_two, b_buffer, b_offset, b_ld);
 
-  // Loads the program from the database
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-
   auto emptyEventList = std::vector<Event>();
   PadCopyTransposeMatrix(queue_, device_, db_, event_, emptyEventList,
                          a_one, a_two, a_ld, a_offset, a_buffer,
                          b_one, b_two, b_ld, b_offset, b_buffer,
-                         alpha, program, false, transpose, conjugate);
+                         alpha, program_, false, transpose, conjugate);
 }
 
 // =================================================================================================