Moved all headers into the source tree, changed headers to .hpp extension

45 files changed, 0 insertions, 2210 deletions

diff --git a/include/internal/routines/common.h b/include/internal/routines/common.h
deleted file mode 100644
index 308785bd..00000000
--- a/include/internal/routines/common.h
+++ /dev/null
@@ -1,173 +0,0 @@
-
-// =================================================================================================
-// 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. This file also contains
-// other common functions to routines, such as a function to launch a kernel.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_COMMON_H_
-#define CLBLAST_ROUTINES_COMMON_H_
-
-#include <string>
-#include <vector>
-
-#include "clblast.h"
-#include "internal/clpp11.h"
-#include "internal/database.h"
-
-namespace clblast {
-// =================================================================================================
-
-// 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);
-
-// =================================================================================================
-
-// 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
deleted file mode 100644
index 42f8f67c..00000000
--- a/include/internal/routines/level1/xamax.h
+++ /dev/null
@@ -1,40 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xamax routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XAMAX_H_
-#define CLBLAST_ROUTINES_XAMAX_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xamax: public Routine {
- public:
-
-  // Constructor
-  Xamax(Queue &queue, EventPointer event, const std::string &name = "AMAX");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoAmax(const size_t n,
-                    const Buffer<unsigned int> &imax_buffer, const size_t imax_offset,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XAMAX_H_
-#endif
diff --git a/include/internal/routines/level1/xasum.h b/include/internal/routines/level1/xasum.h
deleted file mode 100644
index 9d93a6f4..00000000
--- a/include/internal/routines/level1/xasum.h
+++ /dev/null
@@ -1,40 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xasum routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XASUM_H_
-#define CLBLAST_ROUTINES_XASUM_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xasum: public Routine {
- public:
-
-  // Constructor
-  Xasum(Queue &queue, EventPointer event, const std::string &name = "ASUM");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoAsum(const size_t n,
-                    const Buffer<T> &asum_buffer, const size_t asum_offset,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XASUM_H_
-#endif
diff --git a/include/internal/routines/level1/xaxpy.h b/include/internal/routines/level1/xaxpy.h
deleted file mode 100644
index 4c8d2c1f..00000000
--- a/include/internal/routines/level1/xaxpy.h
+++ /dev/null
@@ -1,40 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xaxpy routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XAXPY_H_
-#define CLBLAST_ROUTINES_XAXPY_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xaxpy: public Routine {
- public:
-
-  // Constructor
-  Xaxpy(Queue &queue, EventPointer event, const std::string &name = "AXPY");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoAxpy(const size_t n, const T alpha,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XAXPY_H_
-#endif
diff --git a/include/internal/routines/level1/xcopy.h b/include/internal/routines/level1/xcopy.h
deleted file mode 100644
index c7d03dd0..00000000
--- a/include/internal/routines/level1/xcopy.h
+++ /dev/null
@@ -1,40 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xcopy routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XCOPY_H_
-#define CLBLAST_ROUTINES_XCOPY_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xcopy: public Routine {
- public:
-
-  // Constructor
-  Xcopy(Queue &queue, EventPointer event, const std::string &name = "COPY");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoCopy(const size_t n,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XCOPY_H_
-#endif
diff --git a/include/internal/routines/level1/xdot.h b/include/internal/routines/level1/xdot.h
deleted file mode 100644
index e1968740..00000000
--- a/include/internal/routines/level1/xdot.h
+++ /dev/null
@@ -1,42 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xdot routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XDOT_H_
-#define CLBLAST_ROUTINES_XDOT_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xdot: public Routine {
- public:
-
-  // Constructor
-  Xdot(Queue &queue, EventPointer event, const std::string &name = "DOT");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoDot(const size_t n,
-                   const Buffer<T> &dot_buffer, const size_t dot_offset,
-                   const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                   const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc,
-                   const bool do_conjugate = false);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XDOT_H_
-#endif
diff --git a/include/internal/routines/level1/xdotc.h b/include/internal/routines/level1/xdotc.h
deleted file mode 100644
index 0dc2cfe9..00000000
--- a/include/internal/routines/level1/xdotc.h
+++ /dev/null
@@ -1,44 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xdotc routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XDOTC_H_
-#define CLBLAST_ROUTINES_XDOTC_H_
-
-#include "internal/routines/level1/xdot.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xdotc: public Xdot<T> {
- public:
-
-  // Uses the regular Xdot routine
-  using Xdot<T>::DoDot;
-
-  // Constructor
-  Xdotc(Queue &queue, EventPointer event, const std::string &name = "DOTC");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoDotc(const size_t n,
-                    const Buffer<T> &dot_buffer, const size_t dot_offset,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XDOTC_H_
-#endif
diff --git a/include/internal/routines/level1/xdotu.h b/include/internal/routines/level1/xdotu.h
deleted file mode 100644
index 98988744..00000000
--- a/include/internal/routines/level1/xdotu.h
+++ /dev/null
@@ -1,44 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xdotu routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XDOTU_H_
-#define CLBLAST_ROUTINES_XDOTU_H_
-
-#include "internal/routines/level1/xdot.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xdotu: public Xdot<T> {
- public:
-
-  // Uses the regular Xdot routine
-  using Xdot<T>::DoDot;
-
-  // Constructor
-  Xdotu(Queue &queue, EventPointer event, const std::string &name = "DOTU");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoDotu(const size_t n,
-                    const Buffer<T> &dot_buffer, const size_t dot_offset,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XDOTU_H_
-#endif
diff --git a/include/internal/routines/level1/xmax.h b/include/internal/routines/level1/xmax.h
deleted file mode 100644
index a872cede..00000000
--- a/include/internal/routines/level1/xmax.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xmax routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XMAX_H_
-#define CLBLAST_ROUTINES_XMAX_H_
-
-#include "internal/routine.h"
-#include "internal/routines/level1/xamax.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xmax: public Xamax<T> {
- public:
-
-  // Members and methods from the base class
-  using Xamax<T>::DoAmax;
-
-  // Constructor
-  Xmax(Queue &queue, EventPointer event, const std::string &name = "MAX"):
-    Xamax<T>(queue, event, name) {
-  }
-
-  // Forwards to the regular absolute version. The implementation difference is realised in the
-  // kernel through a pre-processor macro based on the name of the routine.
-  StatusCode DoMax(const size_t n,
-                   const Buffer<unsigned int> &imax_buffer, const size_t imax_offset,
-                   const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc) {
-    return DoAmax(n, imax_buffer, imax_offset, x_buffer, x_offset, x_inc);
-  }
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XMAX_H_
-#endif
diff --git a/include/internal/routines/level1/xmin.h b/include/internal/routines/level1/xmin.h
deleted file mode 100644
index 700c81cc..00000000
--- a/include/internal/routines/level1/xmin.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xmin routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XMIN_H_
-#define CLBLAST_ROUTINES_XMIN_H_
-
-#include "internal/routine.h"
-#include "internal/routines/level1/xamax.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xmin: public Xamax<T> {
- public:
-
-  // Members and methods from the base class
-  using Xamax<T>::DoAmax;
-
-  // Constructor
-  Xmin(Queue &queue, EventPointer event, const std::string &name = "MIN"):
-    Xamax<T>(queue, event, name) {
-  }
-
-  // Forwards to the regular max-absolute version. The implementation difference is realised in the
-  // kernel through a pre-processor macro based on the name of the routine.
-  StatusCode DoMin(const size_t n,
-                   const Buffer<unsigned int> &imin_buffer, const size_t imin_offset,
-                   const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc) {
-    return DoAmax(n, imin_buffer, imin_offset, x_buffer, x_offset, x_inc);
-  }
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XMIN_H_
-#endif
diff --git a/include/internal/routines/level1/xnrm2.h b/include/internal/routines/level1/xnrm2.h
deleted file mode 100644
index ca9268c0..00000000
--- a/include/internal/routines/level1/xnrm2.h
+++ /dev/null
@@ -1,40 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xnrm2 routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XNRM2_H_
-#define CLBLAST_ROUTINES_XNRM2_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xnrm2: public Routine {
- public:
-
-  // Constructor
-  Xnrm2(Queue &queue, EventPointer event, const std::string &name = "NRM2");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoNrm2(const size_t n,
-                    const Buffer<T> &nrm2_buffer, const size_t nrm2_offset,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XNRM2_H_
-#endif
diff --git a/include/internal/routines/level1/xscal.h b/include/internal/routines/level1/xscal.h
deleted file mode 100644
index b9430f3b..00000000
--- a/include/internal/routines/level1/xscal.h
+++ /dev/null
@@ -1,39 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xscal routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSCAL_H_
-#define CLBLAST_ROUTINES_XSCAL_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xscal: public Routine {
- public:
-
-  // Constructor
-  Xscal(Queue &queue, EventPointer event, const std::string &name = "SCAL");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoScal(const size_t n, const T alpha,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSCAL_H_
-#endif
diff --git a/include/internal/routines/level1/xsum.h b/include/internal/routines/level1/xsum.h
deleted file mode 100644
index 2f633b52..00000000
--- a/include/internal/routines/level1/xsum.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xsum routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSUM_H_
-#define CLBLAST_ROUTINES_XSUM_H_
-
-#include "internal/routine.h"
-#include "internal/routines/level1/xasum.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xsum: public Xasum<T> {
- public:
-
-  // Members and methods from the base class
-  using Xasum<T>::DoAsum;
-
-  // Constructor
-  Xsum(Queue &queue, EventPointer event, const std::string &name = "SUM"):
-    Xasum<T>(queue, event, name) {
-  }
-
-  // Forwards to the regular absolute version. The implementation difference is realised in the
-  // kernel through a pre-processor macro based on the name of the routine.
-  StatusCode DoSum(const size_t n,
-                   const Buffer<T> &sum_buffer, const size_t sum_offset,
-                   const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc) {
-    return DoAsum(n, sum_buffer, sum_offset, x_buffer, x_offset, x_inc);
-  }
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSUM_H_
-#endif
diff --git a/include/internal/routines/level1/xswap.h b/include/internal/routines/level1/xswap.h
deleted file mode 100644
index bd063afc..00000000
--- a/include/internal/routines/level1/xswap.h
+++ /dev/null
@@ -1,40 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xswap routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSWAP_H_
-#define CLBLAST_ROUTINES_XSWAP_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xswap: public Routine {
- public:
-
-  // Constructor
-  Xswap(Queue &queue, EventPointer event, const std::string &name = "SWAP");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoSwap(const size_t n,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSWAP_H_
-#endif
diff --git a/include/internal/routines/level2/xgbmv.h b/include/internal/routines/level2/xgbmv.h
deleted file mode 100644
index bc94c77d..00000000
--- a/include/internal/routines/level2/xgbmv.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xgbmv routine. It is based on the generalized mat-vec multiplication
-// routine (Xgemv). The Xgbmv class inherits from the templated class Xgemv, allowing it to call the
-// "MatVec" function directly.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XGBMV_H_
-#define CLBLAST_ROUTINES_XGBMV_H_
-
-#include "internal/routines/level2/xgemv.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xgbmv: public Xgemv<T> {
- public:
-
-  // Uses the generic matrix-vector routine
-  using Xgemv<T>::MatVec;
-
-  // Constructor
-  Xgbmv(Queue &queue, EventPointer event, const std::string &name = "GBMV");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoGbmv(const Layout layout, const Transpose a_transpose,
-                    const size_t m, const size_t n, const size_t kl, const size_t ku,
-                    const T alpha,
-                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const T beta,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XGBMV_H_
-#endif
diff --git a/include/internal/routines/level2/xgemv.h b/include/internal/routines/level2/xgemv.h
deleted file mode 100644
index e9804c62..00000000
--- a/include/internal/routines/level2/xgemv.h
+++ /dev/null
@@ -1,56 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xgemv routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XGEMV_H_
-#define CLBLAST_ROUTINES_XGEMV_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xgemv: public Routine {
- public:
-
-  // Constructor
-  Xgemv(Queue &queue, EventPointer event, const std::string &name = "GEMV");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoGemv(const Layout layout, const Transpose a_transpose,
-                    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> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const T beta,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-
-  // Generic version used also for other matrix-vector multiplications
-  StatusCode MatVec(const Layout layout, const Transpose a_transpose,
-                    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> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const T beta,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc,
-                    bool fast_kernel, bool fast_kernel_rot,
-                    const size_t parameter, const bool packed,
-                    const size_t kl, const size_t ku);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XGEMV_H_
-#endif
diff --git a/include/internal/routines/level2/xger.h b/include/internal/routines/level2/xger.h
deleted file mode 100644
index 184f8477..00000000
--- a/include/internal/routines/level2/xger.h
+++ /dev/null
@@ -1,43 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xger routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XGER_H_
-#define CLBLAST_ROUTINES_XGER_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xger: public Routine {
- public:
-
-  // Constructor
-  Xger(Queue &queue, EventPointer event, const std::string &name = "GER");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoGer(const Layout layout,
-                   const size_t m, const size_t n,
-                   const T alpha,
-                   const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                   const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc,
-                   const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XGER_H_
-#endif
diff --git a/include/internal/routines/level2/xgerc.h b/include/internal/routines/level2/xgerc.h
deleted file mode 100644
index 6d06ef94..00000000
--- a/include/internal/routines/level2/xgerc.h
+++ /dev/null
@@ -1,46 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xgerc routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XGERC_H_
-#define CLBLAST_ROUTINES_XGERC_H_
-
-#include "internal/routines/level2/xger.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xgerc: public Xger<T> {
- public:
-
-  // Uses the regular Xger routine
-  using Xger<T>::DoGer;
-
-  // Constructor
-  Xgerc(Queue &queue, EventPointer event, const std::string &name = "GERC");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoGerc(const Layout layout,
-                    const size_t m, const size_t n,
-                    const T alpha,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc,
-                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XGERC_H_
-#endif
diff --git a/include/internal/routines/level2/xgeru.h b/include/internal/routines/level2/xgeru.h
deleted file mode 100644
index 45ce1cba..00000000
--- a/include/internal/routines/level2/xgeru.h
+++ /dev/null
@@ -1,46 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xgeru routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XGERU_H_
-#define CLBLAST_ROUTINES_XGERU_H_
-
-#include "internal/routines/level2/xger.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xgeru: public Xger<T> {
- public:
-
-  // Uses the regular Xger routine
-  using Xger<T>::DoGer;
-
-  // Constructor
-  Xgeru(Queue &queue, EventPointer event, const std::string &name = "GERU");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoGeru(const Layout layout,
-                    const size_t m, const size_t n,
-                    const T alpha,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc,
-                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XGERU_H_
-#endif
diff --git a/include/internal/routines/level2/xhbmv.h b/include/internal/routines/level2/xhbmv.h
deleted file mode 100644
index f0a6212c..00000000
--- a/include/internal/routines/level2/xhbmv.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xhbmv routine. It is based on the generalized mat-vec multiplication
-// routine (Xgemv). The Xhbmv class inherits from the templated class Xgemv, allowing it to call the
-// "MatVec" function directly.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XHBMV_H_
-#define CLBLAST_ROUTINES_XHBMV_H_
-
-#include "internal/routines/level2/xgemv.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xhbmv: public Xgemv<T> {
- public:
-
-  // Uses the generic matrix-vector routine
-  using Xgemv<T>::MatVec;
-
-  // Constructor
-  Xhbmv(Queue &queue, EventPointer event, const std::string &name = "HBMV");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoHbmv(const Layout layout, const Triangle triangle,
-                    const size_t n, const size_t k,
-                    const T alpha,
-                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const T beta,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XHBMV_H_
-#endif
diff --git a/include/internal/routines/level2/xhemv.h b/include/internal/routines/level2/xhemv.h
deleted file mode 100644
index 3daf2457..00000000
--- a/include/internal/routines/level2/xhemv.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xhemv routine. It is based on the generalized mat-vec multiplication
-// routine (Xgemv). The Xhemv class inherits from the templated class Xgemv, allowing it to call the
-// "MatVec" function directly.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XHEMV_H_
-#define CLBLAST_ROUTINES_XHEMV_H_
-
-#include "internal/routines/level2/xgemv.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xhemv: public Xgemv<T> {
- public:
-
-  // Uses the generic matrix-vector routine
-  using Xgemv<T>::MatVec;
-
-  // Constructor
-  Xhemv(Queue &queue, EventPointer event, const std::string &name = "HEMV");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoHemv(const Layout layout, const Triangle triangle,
-                    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> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const T beta,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XHEMV_H_
-#endif
diff --git a/include/internal/routines/level2/xher.h b/include/internal/routines/level2/xher.h
deleted file mode 100644
index fca8bb97..00000000
--- a/include/internal/routines/level2/xher.h
+++ /dev/null
@@ -1,46 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xher routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XHER_H_
-#define CLBLAST_ROUTINES_XHER_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T, typename U>
-class Xher: public Routine {
- public:
-
-  // Constructor
-  Xher(Queue &queue, EventPointer event, const std::string &name = "HER");
-
-  // Translates alpha of type 'U' into type 'T'
-  T GetAlpha(const U alpha);
-
-  // Templated-precision implementation of the routine
-  StatusCode DoHer(const Layout layout, const Triangle triangle,
-                   const size_t n,
-                   const U alpha,
-                   const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                   const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
-                   const bool packed = false);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XHER_H_
-#endif
diff --git a/include/internal/routines/level2/xher2.h b/include/internal/routines/level2/xher2.h
deleted file mode 100644
index 9a7610f1..00000000
--- a/include/internal/routines/level2/xher2.h
+++ /dev/null
@@ -1,44 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xher2 routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XHER2_H_
-#define CLBLAST_ROUTINES_XHER2_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xher2: public Routine {
- public:
-
-  // Constructor
-  Xher2(Queue &queue, EventPointer event, const std::string &name = "HER2");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoHer2(const Layout layout, const Triangle triangle,
-                    const size_t n,
-                    const T alpha,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc,
-                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const bool packed = false);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XHER2_H_
-#endif
diff --git a/include/internal/routines/level2/xhpmv.h b/include/internal/routines/level2/xhpmv.h
deleted file mode 100644
index a1d5595a..00000000
--- a/include/internal/routines/level2/xhpmv.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xhpmv routine. It is based on the generalized mat-vec multiplication
-// routine (Xgemv). The Xhpmv class inherits from the templated class Xgemv, allowing it to call the
-// "MatVec" function directly.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XHPMV_H_
-#define CLBLAST_ROUTINES_XHPMV_H_
-
-#include "internal/routines/level2/xgemv.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xhpmv: public Xgemv<T> {
- public:
-
-  // Uses the generic matrix-vector routine
-  using Xgemv<T>::MatVec;
-
-  // Constructor
-  Xhpmv(Queue &queue, EventPointer event, const std::string &name = "HPMV");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoHpmv(const Layout layout, const Triangle triangle,
-                    const size_t n,
-                    const T alpha,
-                    const Buffer<T> &ap_buffer, const size_t ap_offset,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const T beta,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XHPMV_H_
-#endif
diff --git a/include/internal/routines/level2/xhpr.h b/include/internal/routines/level2/xhpr.h
deleted file mode 100644
index 6554d74c..00000000
--- a/include/internal/routines/level2/xhpr.h
+++ /dev/null
@@ -1,45 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xhpr routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XHPR_H_
-#define CLBLAST_ROUTINES_XHPR_H_
-
-#include "internal/routines/level2/xher.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T, typename U>
-class Xhpr: public Xher<T,U> {
- public:
-
-  // Uses the regular Xher routine
-  using Xher<T,U>::DoHer;
-
-  // Constructor
-  Xhpr(Queue &queue, EventPointer event, const std::string &name = "HPR");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoHpr(const Layout layout, const Triangle triangle,
-                   const size_t n,
-                   const U alpha,
-                   const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                   const Buffer<T> &ap_buffer, const size_t ap_offset);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XHPR_H_
-#endif
diff --git a/include/internal/routines/level2/xhpr2.h b/include/internal/routines/level2/xhpr2.h
deleted file mode 100644
index d95e7b61..00000000
--- a/include/internal/routines/level2/xhpr2.h
+++ /dev/null
@@ -1,46 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xhpr2 routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XHPR2_H_
-#define CLBLAST_ROUTINES_XHPR2_H_
-
-#include "internal/routines/level2/xher2.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xhpr2: public Xher2<T> {
- public:
-
-  // Uses the regular Xher2 routine
-  using Xher2<T>::DoHer2;
-
-  // Constructor
-  Xhpr2(Queue &queue, EventPointer event, const std::string &name = "HPR2");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoHpr2(const Layout layout, const Triangle triangle,
-                    const size_t n,
-                    const T alpha,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc,
-                    const Buffer<T> &ap_buffer, const size_t ap_offset);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XHPR2_H_
-#endif
diff --git a/include/internal/routines/level2/xsbmv.h b/include/internal/routines/level2/xsbmv.h
deleted file mode 100644
index 4328e377..00000000
--- a/include/internal/routines/level2/xsbmv.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xsbmv routine. It is based on the generalized mat-vec multiplication
-// routine (Xgemv). The Xsbmv class inherits from the templated class Xgemv, allowing it to call the
-// "MatVec" function directly.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSBMV_H_
-#define CLBLAST_ROUTINES_XSBMV_H_
-
-#include "internal/routines/level2/xgemv.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xsbmv: public Xgemv<T> {
- public:
-
-  // Uses the generic matrix-vector routine
-  using Xgemv<T>::MatVec;
-
-  // Constructor
-  Xsbmv(Queue &queue, EventPointer event, const std::string &name = "SBMV");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoSbmv(const Layout layout, const Triangle triangle,
-                    const size_t n, const size_t k,
-                    const T alpha,
-                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const T beta,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSBMV_H_
-#endif
diff --git a/include/internal/routines/level2/xspmv.h b/include/internal/routines/level2/xspmv.h
deleted file mode 100644
index ca3e28b6..00000000
--- a/include/internal/routines/level2/xspmv.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xspmv routine. It is based on the generalized mat-vec multiplication
-// routine (Xgemv). The Xspmv class inherits from the templated class Xgemv, allowing it to call the
-// "MatVec" function directly.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSPMV_H_
-#define CLBLAST_ROUTINES_XSPMV_H_
-
-#include "internal/routines/level2/xgemv.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xspmv: public Xgemv<T> {
- public:
-
-  // Uses the generic matrix-vector routine
-  using Xgemv<T>::MatVec;
-
-  // Constructor
-  Xspmv(Queue &queue, EventPointer event, const std::string &name = "SPMV");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoSpmv(const Layout layout, const Triangle triangle,
-                    const size_t n,
-                    const T alpha,
-                    const Buffer<T> &ap_buffer, const size_t ap_offset,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const T beta,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSPMV_H_
-#endif
diff --git a/include/internal/routines/level2/xspr.h b/include/internal/routines/level2/xspr.h
deleted file mode 100644
index 7e91abc5..00000000
--- a/include/internal/routines/level2/xspr.h
+++ /dev/null
@@ -1,45 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xspr routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSPR_H_
-#define CLBLAST_ROUTINES_XSPR_H_
-
-#include "internal/routines/level2/xher.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xspr: public Xher<T,T> {
- public:
-
-  // Uses the regular Xher routine
-  using Xher<T,T>::DoHer;
-
-  // Constructor
-  Xspr(Queue &queue, EventPointer event, const std::string &name = "SPR");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoSpr(const Layout layout, const Triangle triangle,
-                   const size_t n,
-                   const T alpha,
-                   const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                   const Buffer<T> &ap_buffer, const size_t ap_offset);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSPR_H_
-#endif
diff --git a/include/internal/routines/level2/xspr2.h b/include/internal/routines/level2/xspr2.h
deleted file mode 100644
index a34be8e8..00000000
--- a/include/internal/routines/level2/xspr2.h
+++ /dev/null
@@ -1,46 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xspr2 routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSPR2_H_
-#define CLBLAST_ROUTINES_XSPR2_H_
-
-#include "internal/routines/level2/xher2.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xspr2: public Xher2<T> {
- public:
-
-  // Uses the regular Xher2 routine
-  using Xher2<T>::DoHer2;
-
-  // Constructor
-  Xspr2(Queue &queue, EventPointer event, const std::string &name = "SPR2");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoSpr2(const Layout layout, const Triangle triangle,
-                    const size_t n,
-                    const T alpha,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc,
-                    const Buffer<T> &ap_buffer, const size_t ap_offset);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSPR2_H_
-#endif
diff --git a/include/internal/routines/level2/xsymv.h b/include/internal/routines/level2/xsymv.h
deleted file mode 100644
index 98a0ce88..00000000
--- a/include/internal/routines/level2/xsymv.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xsymv routine. It is based on the generalized mat-vec multiplication
-// routine (Xgemv). The Xsymv class inherits from the templated class Xgemv, allowing it to call the
-// "MatVec" function directly.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSYMV_H_
-#define CLBLAST_ROUTINES_XSYMV_H_
-
-#include "internal/routines/level2/xgemv.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xsymv: public Xgemv<T> {
- public:
-
-  // Uses the generic matrix-vector routine
-  using Xgemv<T>::MatVec;
-
-  // Constructor
-  Xsymv(Queue &queue, EventPointer event, const std::string &name = "SYMV");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoSymv(const Layout layout, const Triangle triangle,
-                    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> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const T beta,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSYMV_H_
-#endif
diff --git a/include/internal/routines/level2/xsyr.h b/include/internal/routines/level2/xsyr.h
deleted file mode 100644
index f88498ae..00000000
--- a/include/internal/routines/level2/xsyr.h
+++ /dev/null
@@ -1,45 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xsyr routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSYR_H_
-#define CLBLAST_ROUTINES_XSYR_H_
-
-#include "internal/routines/level2/xher.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xsyr: public Xher<T,T> {
- public:
-
-  // Uses the regular Xher routine
-  using Xher<T,T>::DoHer;
-
-  // Constructor
-  Xsyr(Queue &queue, EventPointer event, const std::string &name = "SYR");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoSyr(const Layout layout, const Triangle triangle,
-                   const size_t n,
-                   const T alpha,
-                   const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                   const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSYR_H_
-#endif
diff --git a/include/internal/routines/level2/xsyr2.h b/include/internal/routines/level2/xsyr2.h
deleted file mode 100644
index d2d3143a..00000000
--- a/include/internal/routines/level2/xsyr2.h
+++ /dev/null
@@ -1,46 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xsyr2 routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSYR2_H_
-#define CLBLAST_ROUTINES_XSYR2_H_
-
-#include "internal/routines/level2/xher2.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xsyr2: public Xher2<T> {
- public:
-
-  // Uses the regular Xher2 routine
-  using Xher2<T>::DoHer2;
-
-  // Constructor
-  Xsyr2(Queue &queue, EventPointer event, const std::string &name = "SYR2");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoSyr2(const Layout layout, const Triangle triangle,
-                    const size_t n,
-                    const T alpha,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc,
-                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSYR2_H_
-#endif
diff --git a/include/internal/routines/level2/xtbmv.h b/include/internal/routines/level2/xtbmv.h
deleted file mode 100644
index 493a9853..00000000
--- a/include/internal/routines/level2/xtbmv.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xtbmv routine. It is based on the generalized mat-vec multiplication
-// routine (Xgemv). The Xtbmv class inherits from the templated class Xgemv, allowing it to call the
-// "MatVec" function directly.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XTBMV_H_
-#define CLBLAST_ROUTINES_XTBMV_H_
-
-#include "internal/routines/level2/xgemv.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xtbmv: public Xgemv<T> {
- public:
-
-  // Uses the generic matrix-vector routine
-  using Xgemv<T>::queue_;
-  using Xgemv<T>::context_;
-  using Xgemv<T>::MatVec;
-
-  // Constructor
-  Xtbmv(Queue &queue, EventPointer event, const std::string &name = "TBMV");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoTbmv(const Layout layout, const Triangle triangle,
-                    const Transpose a_transpose, const Diagonal diagonal,
-                    const size_t n, const size_t k,
-                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XTBMV_H_
-#endif
diff --git a/include/internal/routines/level2/xtpmv.h b/include/internal/routines/level2/xtpmv.h
deleted file mode 100644
index ce5cae6f..00000000
--- a/include/internal/routines/level2/xtpmv.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xtpmv routine. It is based on the generalized mat-vec multiplication
-// routine (Xgemv). The Xtpmv class inherits from the templated class Xgemv, allowing it to call the
-// "MatVec" function directly.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XTPMV_H_
-#define CLBLAST_ROUTINES_XTPMV_H_
-
-#include "internal/routines/level2/xgemv.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xtpmv: public Xgemv<T> {
- public:
-
-  // Uses the generic matrix-vector routine
-  using Xgemv<T>::queue_;
-  using Xgemv<T>::context_;
-  using Xgemv<T>::MatVec;
-
-  // Constructor
-  Xtpmv(Queue &queue, EventPointer event, const std::string &name = "TPMV");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoTpmv(const Layout layout, const Triangle triangle,
-                    const Transpose a_transpose, const Diagonal diagonal,
-                    const size_t n,
-                    const Buffer<T> &ap_buffer, const size_t ap_offset,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XTPMV_H_
-#endif
diff --git a/include/internal/routines/level2/xtrmv.h b/include/internal/routines/level2/xtrmv.h
deleted file mode 100644
index 4407bad7..00000000
--- a/include/internal/routines/level2/xtrmv.h
+++ /dev/null
@@ -1,49 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xtrmv routine. It is based on the generalized mat-vec multiplication
-// routine (Xgemv). The Xtrmv class inherits from the templated class Xgemv, allowing it to call the
-// "MatVec" function directly.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XTRMV_H_
-#define CLBLAST_ROUTINES_XTRMV_H_
-
-#include "internal/routines/level2/xgemv.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xtrmv: public Xgemv<T> {
- public:
-
-  // Uses the generic matrix-vector routine
-  using Xgemv<T>::queue_;
-  using Xgemv<T>::context_;
-  using Xgemv<T>::MatVec;
-
-  // Constructor
-  Xtrmv(Queue &queue, EventPointer event, const std::string &name = "TRMV");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoTrmv(const Layout layout, const Triangle triangle,
-                    const Transpose a_transpose, const Diagonal diagonal,
-                    const size_t n,
-                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XTRMV_H_
-#endif
diff --git a/include/internal/routines/level3/xgemm.h b/include/internal/routines/level3/xgemm.h
deleted file mode 100644
index c0541eef..00000000
--- a/include/internal/routines/level3/xgemm.h
+++ /dev/null
@@ -1,48 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xgemm routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XGEMM_H_
-#define CLBLAST_ROUTINES_XGEMM_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xgemm: public Routine {
- public:
-
-  // Constructor
-  Xgemm(Queue &queue, EventPointer event, const std::string &name = "GEMM");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoGemm(const Layout layout, const Transpose a_transpose, const Transpose b_transpose,
-                    const size_t m, const size_t n, const size_t k,
-                    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);
-
- protected:
-  // Static variable to get the precision
-  const static Precision precision_;
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XGEMM_H_
-#endif
diff --git a/include/internal/routines/level3/xhemm.h b/include/internal/routines/level3/xhemm.h
deleted file mode 100644
index e0f35669..00000000
--- a/include/internal/routines/level3/xhemm.h
+++ /dev/null
@@ -1,54 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xhemm routine. It is based on the generalized matrix multiplication
-// routine (Xgemm). The implementation is very similar to the Xsymm routine.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XHEMM_H_
-#define CLBLAST_ROUTINES_XHEMM_H_
-
-#include "internal/routines/level3/xgemm.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xhemm: public Xgemm<T> {
- public:
-
-  // Uses methods and variables the regular Xgemm routine
-  using Xgemm<T>::precision_;
-  using Xgemm<T>::routine_name_;
-  using Xgemm<T>::queue_;
-  using Xgemm<T>::context_;
-  using Xgemm<T>::device_;
-  using Xgemm<T>::db_;
-  using Xgemm<T>::DoGemm;
-
-  // Constructor
-  Xhemm(Queue &queue, EventPointer event, const std::string &name = "HEMM");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoHemm(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);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XHEMM_H_
-#endif
diff --git a/include/internal/routines/level3/xher2k.h b/include/internal/routines/level3/xher2k.h
deleted file mode 100644
index b7764e18..00000000
--- a/include/internal/routines/level3/xher2k.h
+++ /dev/null
@@ -1,46 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xher2k routine. The precision is implemented using the template argument
-// 'T', whereas the alpha/beta arguments are of type 'U'. The implementation is very similar to the
-// Xsyr2k routine.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XHER2K_H_
-#define CLBLAST_ROUTINES_XHER2K_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T, typename U>
-class Xher2k: public Routine {
- public:
-
-  // Constructor
-  Xher2k(Queue &queue, EventPointer event, const std::string &name = "HER2K");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoHer2k(const Layout layout, const Triangle triangle, const Transpose ab_transpose,
-                     const size_t n, const size_t k,
-                     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 U beta,
-                     const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XHER2K_H_
-#endif
diff --git a/include/internal/routines/level3/xherk.h b/include/internal/routines/level3/xherk.h
deleted file mode 100644
index abcf4c1a..00000000
--- a/include/internal/routines/level3/xherk.h
+++ /dev/null
@@ -1,45 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xherk routine. The precision is implemented using the template argument
-// 'T', whereas the alpha/beta arguments are of type 'U'. The implementation is very similar to the
-// Xsyrk routine.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XHERK_H_
-#define CLBLAST_ROUTINES_XHERK_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T, typename U>
-class Xherk: public Routine {
- public:
-
-  // Constructor
-  Xherk(Queue &queue, EventPointer event, const std::string &name = "HERK");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoHerk(const Layout layout, const Triangle triangle, const Transpose a_transpose,
-                    const size_t n, const size_t k,
-                    const U alpha,
-                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const U beta,
-                    const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XHERK_H_
-#endif
diff --git a/include/internal/routines/level3/xsymm.h b/include/internal/routines/level3/xsymm.h
deleted file mode 100644
index 889abfb7..00000000
--- a/include/internal/routines/level3/xsymm.h
+++ /dev/null
@@ -1,56 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xsymm routine. It is based on the generalized matrix multiplication
-// routine (Xgemm). The Xsymm class inherits from the templated class Xgemm, allowing it to call the
-// "DoGemm" function directly. The "DoSymm" function first preprocesses the symmetric matrix by
-// transforming it into a general matrix, and then calls the regular GEMM code.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSYMM_H_
-#define CLBLAST_ROUTINES_XSYMM_H_
-
-#include "internal/routines/level3/xgemm.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xsymm: public Xgemm<T> {
- public:
-
-  // Uses methods and variables the regular Xgemm routine
-  using Xgemm<T>::precision_;
-  using Xgemm<T>::routine_name_;
-  using Xgemm<T>::queue_;
-  using Xgemm<T>::context_;
-  using Xgemm<T>::device_;
-  using Xgemm<T>::db_;
-  using Xgemm<T>::DoGemm;
-
-  // Constructor
-  Xsymm(Queue &queue, EventPointer event, const std::string &name = "SYMM");
-
-  // Templated-precision implementation of the routine
-  StatusCode 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);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSYMM_H_
-#endif
diff --git a/include/internal/routines/level3/xsyr2k.h b/include/internal/routines/level3/xsyr2k.h
deleted file mode 100644
index f75c91e5..00000000
--- a/include/internal/routines/level3/xsyr2k.h
+++ /dev/null
@@ -1,46 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xsyr2k routine. The precision is implemented using a template argument.
-// The implementation is very similar to Xsyrk (see header for details), except for the fact that
-// the main XgemmUpper/XgemmLower kernel is called twice: C = AB^T + C and C = BA^T + C.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSYR2K_H_
-#define CLBLAST_ROUTINES_XSYR2K_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xsyr2k: public Routine {
- public:
-
-  // Constructor
-  Xsyr2k(Queue &queue, EventPointer event, const std::string &name = "SYR2K");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoSyr2k(const Layout layout, const Triangle triangle, const Transpose ab_transpose,
-                     const size_t n, const size_t k,
-                     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);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSYR2K_H_
-#endif
diff --git a/include/internal/routines/level3/xsyrk.h b/include/internal/routines/level3/xsyrk.h
deleted file mode 100644
index 0710fa74..00000000
--- a/include/internal/routines/level3/xsyrk.h
+++ /dev/null
@@ -1,47 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xsyrk routine. The precision is implemented using a template argument.
-// The implementation is based on the regular Xgemm routine and kernel, but with two main changes:
-// 1) The final unpad(transpose) kernel updates only the upper/lower triangular part.
-// 2) The main Xgemm kernel masks workgroups not contributing to usefull data. This is only for
-//    performance reasons, as the actual masking is done later (see the first point).
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XSYRK_H_
-#define CLBLAST_ROUTINES_XSYRK_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xsyrk: public Routine {
- public:
-
-  // Constructor
-  Xsyrk(Queue &queue, EventPointer event, const std::string &name = "SYRK");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoSyrk(const Layout layout, const Triangle triangle, const Transpose a_transpose,
-                    const size_t n, const size_t k,
-                    const T alpha,
-                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const T beta,
-                    const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XSYRK_H_
-#endif
diff --git a/include/internal/routines/level3/xtrmm.h b/include/internal/routines/level3/xtrmm.h
deleted file mode 100644
index e18ad17a..00000000
--- a/include/internal/routines/level3/xtrmm.h
+++ /dev/null
@@ -1,54 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xtrmm routine. The implementation is based on first transforming the
-// upper/lower unit/non-unit triangular matrix into a regular matrix and then calling the GEMM
-// routine. Therefore, this class inherits from the Xgemm class.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XTRMM_H_
-#define CLBLAST_ROUTINES_XTRMM_H_
-
-#include "internal/routines/level3/xgemm.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xtrmm: public Xgemm<T> {
- public:
-
-  // Uses methods and variables the regular Xgemm routine
-  using Xgemm<T>::precision_;
-  using Xgemm<T>::routine_name_;
-  using Xgemm<T>::queue_;
-  using Xgemm<T>::context_;
-  using Xgemm<T>::device_;
-  using Xgemm<T>::db_;
-  using Xgemm<T>::DoGemm;
-
-  // Constructor
-  Xtrmm(Queue &queue, EventPointer event, const std::string &name = "TRMM");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoTrmm(const Layout layout, const Side side, const Triangle triangle,
-                    const Transpose a_transpose, const Diagonal diagonal,
-                    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);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XTRMM_H_
-#endif
diff --git a/include/internal/routines/levelx/xomatcopy.h b/include/internal/routines/levelx/xomatcopy.h
deleted file mode 100644
index d2acb50d..00000000
--- a/include/internal/routines/levelx/xomatcopy.h
+++ /dev/null
@@ -1,41 +0,0 @@
-
-// =================================================================================================
-// 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 implements the Xomatcopy routine. The precision is implemented using a template argument.
-//
-// =================================================================================================
-
-#ifndef CLBLAST_ROUTINES_XOMATCOPY_H_
-#define CLBLAST_ROUTINES_XOMATCOPY_H_
-
-#include "internal/routine.h"
-
-namespace clblast {
-// =================================================================================================
-
-// See comment at top of file for a description of the class
-template <typename T>
-class Xomatcopy: public Routine {
- public:
-
-  // Constructor
-  Xomatcopy(Queue &queue, EventPointer event, const std::string &name = "OMATCOPY");
-
-  // Templated-precision implementation of the routine
-  StatusCode DoOmatcopy(const Layout layout, const Transpose a_transpose,
-                        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);
-};
-
-// =================================================================================================
-} // namespace clblast
-
-// CLBLAST_ROUTINES_XOMATCOPY_H_
-#endif