Moved the RunKernel and PadCopyTransposeMatrix functions out of the Routine class

26 files changed, 204 insertions, 53 deletions

diff --git a/include/internal/routine.h b/include/internal/routine.h
index 0f64c479..e1888f1f 100644
--- a/include/internal/routine.h
+++ b/include/internal/routine.h
@@ -40,30 +40,6 @@ class Routine {
   StatusCode SetUp();
 
  protected:
-  
-  // Runs a kernel given the global and local thread sizes
-  StatusCode RunKernel(Kernel &kernel, std::vector<size_t> global,
-                       const std::vector<size_t> &local, EventPointer event,
-                       std::vector<Event>& waitForEvents);
-
-  // As above, but without an event waiting list
-  StatusCode RunKernel(Kernel &kernel, std::vector<size_t> global,
-                       const std::vector<size_t> &local, EventPointer event);
-
-  // Copies/transposes a matrix and padds/unpads it with zeroes. This method is also able to write
-  // to symmetric and triangular matrices through optional arguments.
-  StatusCode PadCopyTransposeMatrix(EventPointer event, std::vector<Event>& waitForEvents,
-                                    const size_t src_one, const size_t src_two,
-                                    const size_t src_ld, const size_t src_offset,
-                                    const Buffer<T> &src,
-                                    const size_t dest_one, const size_t dest_two,
-                                    const size_t dest_ld, const size_t dest_offset,
-                                    const Buffer<T> &dest,
-                                    const T alpha,
-                                    const Program &program, const bool do_pad,
-                                    const bool do_transpose, const bool do_conjugate,
-                                    const bool upper = false, const bool lower = false,
-                                    const bool diagonal_imag_zero = false);
 
   // Stores a newly compiled binary/program into the cache
   void StoreBinaryToCache(const std::string& binary) const {
@@ -105,16 +81,28 @@ class Routine {
 
   // OpenCL device properties
   const std::string device_name_;
-  const size_t max_work_item_dimensions_;
-  const std::vector<size_t> max_work_item_sizes_;
-  const size_t max_work_group_size_;
 
   // Connection to the database for all the device-specific parameters
   const Database db_;
 };
 
 // =================================================================================================
+
+// Enqueues a kernel, waits for completion, and checks for errors
+StatusCode RunKernel(Kernel &kernel, Queue queue, const Device device,
+                     std::vector<size_t> global, const std::vector<size_t> &local,
+                     EventPointer event, std::vector<Event>& waitForEvents);
+
+// As above, but without an event waiting list
+StatusCode RunKernel(Kernel &kernel, Queue queue, const Device device,
+                     std::vector<size_t> global, const std::vector<size_t> &local,
+                     EventPointer event);
+
+// =================================================================================================
 } // namespace clblast
 
+// Temporary fix: TODO place include in a more logical place
+#include "internal/routines/common.h"
+
 // CLBLAST_ROUTINE_H_
 #endif
diff --git a/include/internal/routines/common.h b/include/internal/routines/common.h
new file mode 100644
index 00000000..95fbde46
--- /dev/null
+++ b/include/internal/routines/common.h
@@ -0,0 +1,159 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file contains all the interfaces to common kernels, such as copying, padding, and
+// transposing a matrix. These functions are templated and thus header-only.
+//
+// =================================================================================================
+
+#ifndef CLBLAST_ROUTINES_COMMON_H_
+#define CLBLAST_ROUTINES_COMMON_H_
+
+#include <string>
+#include <vector>
+
+#include "internal/utilities.h"
+#include "internal/routine.h"
+
+namespace clblast {
+// =================================================================================================
+
+// Copies or transposes a matrix and optionally pads/unpads it with zeros. This method is also able
+// to write to symmetric and triangular matrices through optional arguments.
+template <typename T>
+StatusCode PadCopyTransposeMatrix(Queue queue, const Device device, const Context context,
+                                  const Database db,
+                                  EventPointer event, std::vector<Event>& waitForEvents,
+                                  const size_t src_one, const size_t src_two,
+                                  const size_t src_ld, const size_t src_offset,
+                                  const Buffer<T> &src,
+                                  const size_t dest_one, const size_t dest_two,
+                                  const size_t dest_ld, const size_t dest_offset,
+                                  const Buffer<T> &dest,
+                                  const T alpha,
+                                  const Program &program, const bool do_pad,
+                                  const bool do_transpose, const bool do_conjugate,
+                                  const bool upper = false, const bool lower = false,
+                                  const bool diagonal_imag_zero = false) {
+
+  // Determines whether or not the fast-version could potentially be used
+  auto use_fast_kernel = (src_offset == 0) && (dest_offset == 0) && (do_conjugate == false) &&
+                         (src_one == dest_one) && (src_two == dest_two) && (src_ld == dest_ld) &&
+                         (upper == false) && (lower == false) && (diagonal_imag_zero == false);
+
+  // Determines the right kernel
+  auto kernel_name = std::string{};
+  if (do_transpose) {
+    if (use_fast_kernel &&
+        IsMultiple(src_ld, db["TRA_WPT"]) &&
+        IsMultiple(src_one, db["TRA_WPT"]*db["TRA_WPT"]) &&
+        IsMultiple(src_two, db["TRA_WPT"]*db["TRA_WPT"])) {
+      kernel_name = "TransposeMatrixFast";
+    }
+    else {
+      use_fast_kernel = false;
+      kernel_name = (do_pad) ? "TransposePadMatrix" : "TransposeMatrix";
+    }
+  }
+  else {
+    if (use_fast_kernel &&
+        IsMultiple(src_ld, db["COPY_VW"]) &&
+        IsMultiple(src_one, db["COPY_VW"]*db["COPY_DIMX"]) &&
+        IsMultiple(src_two, db["COPY_WPT"]*db["COPY_DIMY"])) {
+      kernel_name = "CopyMatrixFast";
+    }
+    else {
+      use_fast_kernel = false;
+      kernel_name = (do_pad) ? "CopyPadMatrix" : "CopyMatrix";
+    }
+  }
+
+  // Upload the scalar argument as a constant buffer to the device (needed for half-precision)
+  auto alpha_buffer = Buffer<T>(context, 1);
+  alpha_buffer.Write(queue, 1, &alpha);
+
+  // Retrieves the kernel from the compiled binary
+  try {
+    auto kernel = Kernel(program, kernel_name);
+
+    // Sets the kernel arguments
+    if (use_fast_kernel) {
+      kernel.SetArgument(0, static_cast<int>(src_ld));
+      kernel.SetArgument(1, src());
+      kernel.SetArgument(2, dest());
+      kernel.SetArgument(3, alpha_buffer());
+    }
+    else {
+      kernel.SetArgument(0, static_cast<int>(src_one));
+      kernel.SetArgument(1, static_cast<int>(src_two));
+      kernel.SetArgument(2, static_cast<int>(src_ld));
+      kernel.SetArgument(3, static_cast<int>(src_offset));
+      kernel.SetArgument(4, src());
+      kernel.SetArgument(5, static_cast<int>(dest_one));
+      kernel.SetArgument(6, static_cast<int>(dest_two));
+      kernel.SetArgument(7, static_cast<int>(dest_ld));
+      kernel.SetArgument(8, static_cast<int>(dest_offset));
+      kernel.SetArgument(9, dest());
+      kernel.SetArgument(10, alpha_buffer());
+      if (do_pad) {
+        kernel.SetArgument(11, static_cast<int>(do_conjugate));
+      }
+      else {
+        kernel.SetArgument(11, static_cast<int>(upper));
+        kernel.SetArgument(12, static_cast<int>(lower));
+        kernel.SetArgument(13, static_cast<int>(diagonal_imag_zero));
+      }
+    }
+
+    // Launches the kernel and returns the error code. Uses global and local thread sizes based on
+    // parameters in the database.
+    if (do_transpose) {
+      if (use_fast_kernel) {
+        const auto global = std::vector<size_t>{
+          dest_one / db["TRA_WPT"],
+          dest_two / db["TRA_WPT"]
+        };
+        const auto local = std::vector<size_t>{db["TRA_DIM"], db["TRA_DIM"]};
+        return RunKernel(kernel, queue, device, global, local, event, waitForEvents);
+      }
+      else {
+        const auto global = std::vector<size_t>{
+          Ceil(CeilDiv(dest_one, db["PADTRA_WPT"]), db["PADTRA_TILE"]),
+          Ceil(CeilDiv(dest_two, db["PADTRA_WPT"]), db["PADTRA_TILE"])
+        };
+        const auto local = std::vector<size_t>{db["PADTRA_TILE"], db["PADTRA_TILE"]};
+        return RunKernel(kernel, queue, device, global, local, event, waitForEvents);
+      }
+    }
+    else {
+      if (use_fast_kernel) {
+        const auto global = std::vector<size_t>{
+          dest_one / db["COPY_VW"],
+          dest_two / db["COPY_WPT"]
+        };
+        const auto local = std::vector<size_t>{db["COPY_DIMX"], db["COPY_DIMY"]};
+        return RunKernel(kernel, queue, device, global, local, event, waitForEvents);
+      }
+      else {
+        const auto global = std::vector<size_t>{
+          Ceil(CeilDiv(dest_one, db["PAD_WPTX"]), db["PAD_DIMX"]),
+          Ceil(CeilDiv(dest_two, db["PAD_WPTY"]), db["PAD_DIMY"])
+        };
+        const auto local = std::vector<size_t>{db["PAD_DIMX"], db["PAD_DIMY"]};
+        return RunKernel(kernel, queue, device, global, local, event, waitForEvents);
+      }
+    }
+  } catch (...) { return StatusCode::kInvalidKernel; }
+}
+
+// =================================================================================================
+} // namespace clblast
+
+// CLBLAST_ROUTINES_COMMON_H_
+#endif
diff --git a/include/internal/routines/level1/xamax.h b/include/internal/routines/level1/xamax.h
index 70d8a6b0..54434362 100644
--- a/include/internal/routines/level1/xamax.h
+++ b/include/internal/routines/level1/xamax.h
@@ -28,10 +28,10 @@ class Xamax: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xamax(Queue &queue, EventPointer event, const std::string &name = "AMAX");
diff --git a/include/internal/routines/level1/xasum.h b/include/internal/routines/level1/xasum.h
index 8b5c9c76..ee593e30 100644
--- a/include/internal/routines/level1/xasum.h
+++ b/include/internal/routines/level1/xasum.h
@@ -28,10 +28,10 @@ class Xasum: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xasum(Queue &queue, EventPointer event, const std::string &name = "ASUM");
diff --git a/include/internal/routines/level1/xaxpy.h b/include/internal/routines/level1/xaxpy.h
index bd8f5892..6ea3264b 100644
--- a/include/internal/routines/level1/xaxpy.h
+++ b/include/internal/routines/level1/xaxpy.h
@@ -28,10 +28,10 @@ class Xaxpy: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xaxpy(Queue &queue, EventPointer event, const std::string &name = "AXPY");
diff --git a/include/internal/routines/level1/xcopy.h b/include/internal/routines/level1/xcopy.h
index f01d5b08..b371ca9b 100644
--- a/include/internal/routines/level1/xcopy.h
+++ b/include/internal/routines/level1/xcopy.h
@@ -28,9 +28,9 @@ class Xcopy: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xcopy(Queue &queue, EventPointer event, const std::string &name = "COPY");
diff --git a/include/internal/routines/level1/xdot.h b/include/internal/routines/level1/xdot.h
index 46cf9959..7c69a902 100644
--- a/include/internal/routines/level1/xdot.h
+++ b/include/internal/routines/level1/xdot.h
@@ -28,10 +28,10 @@ class Xdot: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xdot(Queue &queue, EventPointer event, const std::string &name = "DOT");
diff --git a/include/internal/routines/level1/xnrm2.h b/include/internal/routines/level1/xnrm2.h
index 3a0cf45c..f83cc2ce 100644
--- a/include/internal/routines/level1/xnrm2.h
+++ b/include/internal/routines/level1/xnrm2.h
@@ -28,10 +28,10 @@ class Xnrm2: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xnrm2(Queue &queue, EventPointer event, const std::string &name = "NRM2");
diff --git a/include/internal/routines/level1/xscal.h b/include/internal/routines/level1/xscal.h
index 9a0f83ab..40f017f2 100644
--- a/include/internal/routines/level1/xscal.h
+++ b/include/internal/routines/level1/xscal.h
@@ -28,9 +28,9 @@ class Xscal: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xscal(Queue &queue, EventPointer event, const std::string &name = "SCAL");
diff --git a/include/internal/routines/level1/xswap.h b/include/internal/routines/level1/xswap.h
index 02360c38..f794a1b4 100644
--- a/include/internal/routines/level1/xswap.h
+++ b/include/internal/routines/level1/xswap.h
@@ -28,9 +28,9 @@ class Xswap: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xswap(Queue &queue, EventPointer event, const std::string &name = "SWAP");
diff --git a/include/internal/routines/level2/xgemv.h b/include/internal/routines/level2/xgemv.h
index dcfcbd1f..aec8b35b 100644
--- a/include/internal/routines/level2/xgemv.h
+++ b/include/internal/routines/level2/xgemv.h
@@ -28,10 +28,10 @@ class Xgemv: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xgemv(Queue &queue, EventPointer event, const std::string &name = "GEMV");
diff --git a/include/internal/routines/level2/xger.h b/include/internal/routines/level2/xger.h
index a26ca24a..260325cb 100644
--- a/include/internal/routines/level2/xger.h
+++ b/include/internal/routines/level2/xger.h
@@ -28,10 +28,10 @@ class Xger: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xger(Queue &queue, EventPointer event, const std::string &name = "GER");
diff --git a/include/internal/routines/level2/xher.h b/include/internal/routines/level2/xher.h
index d32d337e..d66b2603 100644
--- a/include/internal/routines/level2/xher.h
+++ b/include/internal/routines/level2/xher.h
@@ -28,10 +28,10 @@ class Xher: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xher(Queue &queue, EventPointer event, const std::string &name = "HER");
diff --git a/include/internal/routines/level2/xher2.h b/include/internal/routines/level2/xher2.h
index 956ffe6a..35bf8190 100644
--- a/include/internal/routines/level2/xher2.h
+++ b/include/internal/routines/level2/xher2.h
@@ -28,10 +28,10 @@ class Xher2: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xher2(Queue &queue, EventPointer event, const std::string &name = "HER2");
diff --git a/include/internal/routines/level2/xtbmv.h b/include/internal/routines/level2/xtbmv.h
index 3b358080..c9107c25 100644
--- a/include/internal/routines/level2/xtbmv.h
+++ b/include/internal/routines/level2/xtbmv.h
@@ -28,6 +28,7 @@ class Xtbmv: public Xgemv<T> {
   
   // Members from the base class
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::context_;
 
   // Uses the generic matrix-vector routine
diff --git a/include/internal/routines/level2/xtpmv.h b/include/internal/routines/level2/xtpmv.h
index f306cf4a..e85c225f 100644
--- a/include/internal/routines/level2/xtpmv.h
+++ b/include/internal/routines/level2/xtpmv.h
@@ -28,6 +28,7 @@ class Xtpmv: public Xgemv<T> {
   
   // Members from the base class
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::context_;
 
   // Uses the generic matrix-vector routine
diff --git a/include/internal/routines/level2/xtrmv.h b/include/internal/routines/level2/xtrmv.h
index cf0824a4..97a180ff 100644
--- a/include/internal/routines/level2/xtrmv.h
+++ b/include/internal/routines/level2/xtrmv.h
@@ -28,6 +28,7 @@ class Xtrmv: public Xgemv<T> {
   
   // Members from the base class
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::context_;
 
   // Uses the generic matrix-vector routine
diff --git a/include/internal/routines/level3/xgemm.h b/include/internal/routines/level3/xgemm.h
index 8facaa76..2fd853a9 100644
--- a/include/internal/routines/level3/xgemm.h
+++ b/include/internal/routines/level3/xgemm.h
@@ -28,11 +28,10 @@ class Xgemm: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::PadCopyTransposeMatrix;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xgemm(Queue &queue, EventPointer event, const std::string &name = "GEMM");
diff --git a/include/internal/routines/level3/xhemm.h b/include/internal/routines/level3/xhemm.h
index cf833f57..8bd38393 100644
--- a/include/internal/routines/level3/xhemm.h
+++ b/include/internal/routines/level3/xhemm.h
@@ -27,9 +27,10 @@ class Xhemm: public Xgemm<T> {
 
   // Members and methods from the base class
   using Routine<T>::db_;
+  using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Uses the regular Xgemm routine
   using Xgemm<T>::DoGemm;
diff --git a/include/internal/routines/level3/xher2k.h b/include/internal/routines/level3/xher2k.h
index cdba33ab..1afe87a6 100644
--- a/include/internal/routines/level3/xher2k.h
+++ b/include/internal/routines/level3/xher2k.h
@@ -30,11 +30,10 @@ class Xher2k: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::PadCopyTransposeMatrix;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xher2k(Queue &queue, EventPointer event, const std::string &name = "HER2K");
diff --git a/include/internal/routines/level3/xherk.h b/include/internal/routines/level3/xherk.h
index a9d1615a..64abae3b 100644
--- a/include/internal/routines/level3/xherk.h
+++ b/include/internal/routines/level3/xherk.h
@@ -30,11 +30,10 @@ class Xherk: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::PadCopyTransposeMatrix;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xherk(Queue &queue, EventPointer event, const std::string &name = "HERK");
diff --git a/include/internal/routines/level3/xsymm.h b/include/internal/routines/level3/xsymm.h
index ec145a89..c35dfb5e 100644
--- a/include/internal/routines/level3/xsymm.h
+++ b/include/internal/routines/level3/xsymm.h
@@ -29,9 +29,10 @@ class Xsymm: public Xgemm<T> {
 
   // Members and methods from the base class
   using Routine<T>::db_;
+  using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Uses the regular Xgemm routine
   using Xgemm<T>::DoGemm;
diff --git a/include/internal/routines/level3/xsyr2k.h b/include/internal/routines/level3/xsyr2k.h
index 1093d635..73d11b0b 100644
--- a/include/internal/routines/level3/xsyr2k.h
+++ b/include/internal/routines/level3/xsyr2k.h
@@ -30,11 +30,10 @@ class Xsyr2k: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::PadCopyTransposeMatrix;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xsyr2k(Queue &queue, EventPointer event, const std::string &name = "SYR2K");
diff --git a/include/internal/routines/level3/xsyrk.h b/include/internal/routines/level3/xsyrk.h
index c2edbf2b..344c02e2 100644
--- a/include/internal/routines/level3/xsyrk.h
+++ b/include/internal/routines/level3/xsyrk.h
@@ -32,11 +32,10 @@ class Xsyrk: public Routine<T> {
   using Routine<T>::db_;
   using Routine<T>::source_string_;
   using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::PadCopyTransposeMatrix;
-  using Routine<T>::RunKernel;
 
   // Constructor
   Xsyrk(Queue &queue, EventPointer event, const std::string &name = "SYRK");
diff --git a/include/internal/routines/level3/xtrmm.h b/include/internal/routines/level3/xtrmm.h
index 6cb46b01..5c12815d 100644
--- a/include/internal/routines/level3/xtrmm.h
+++ b/include/internal/routines/level3/xtrmm.h
@@ -28,9 +28,10 @@ class Xtrmm: public Xgemm<T> {
 
   // Members and methods from the base class
   using Routine<T>::db_;
+  using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::RunKernel;
 
   // Uses the regular Xgemm routine
   using Xgemm<T>::DoGemm;
diff --git a/include/internal/routines/levelx/xomatcopy.h b/include/internal/routines/levelx/xomatcopy.h
index ee38fa15..7c284635 100644
--- a/include/internal/routines/levelx/xomatcopy.h
+++ b/include/internal/routines/levelx/xomatcopy.h
@@ -25,10 +25,13 @@ class Xomatcopy: public Routine<T> {
  public:
 
   // Members and methods from the base class
+  using Routine<T>::db_;
   using Routine<T>::source_string_;
+  using Routine<T>::queue_;
+  using Routine<T>::device_;
   using Routine<T>::event_;
+  using Routine<T>::context_;
   using Routine<T>::GetProgramFromCache;
-  using Routine<T>::PadCopyTransposeMatrix;
 
   // Constructor
   Xomatcopy(Queue &queue, EventPointer event, const std::string &name = "OMATCOPY");