Now using the new Claduc C++11 OpenCL header

10 files changed, 162 insertions, 70 deletions

diff --git a/include/internal/routines/level1/xaxpy.h b/include/internal/routines/level1/xaxpy.h
index e548e553..4b9da890 100644
--- a/include/internal/routines/level1/xaxpy.h
+++ b/include/internal/routines/level1/xaxpy.h
@@ -21,14 +21,26 @@ namespace clblast {
 
 // See comment at top of file for a description of the class
 template <typename T>
-class Xaxpy: public Routine {
+class Xaxpy: public Routine<T> {
  public:
-  Xaxpy(CommandQueue &queue, Event &event);
+
+  // Members and methods from the base class
+  using Routine<T>::db_;
+  using Routine<T>::source_string_;
+  using Routine<T>::queue_;
+  using Routine<T>::GetProgramFromCache;
+  using Routine<T>::TestVectorX;
+  using Routine<T>::TestVectorY;
+  using Routine<T>::RunKernel;
+  using Routine<T>::ErrorIn;
+
+  // Constructor
+  Xaxpy(Queue &queue, Event &event);
 
   // Templated-precision implementation of the routine
   StatusCode DoAxpy(const size_t n, const T alpha,
-                    const Buffer &x_buffer, const size_t x_offset, const size_t x_inc,
-                    const Buffer &y_buffer, const size_t y_offset, const size_t y_inc);
+                    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);
 
  private:
   // Static variable to get the precision
diff --git a/include/internal/routines/level2/xgemv.h b/include/internal/routines/level2/xgemv.h
index a3109036..5ada9b03 100644
--- a/include/internal/routines/level2/xgemv.h
+++ b/include/internal/routines/level2/xgemv.h
@@ -21,18 +21,31 @@ namespace clblast {
 
 // See comment at top of file for a description of the class
 template <typename T>
-class Xgemv: public Routine {
+class Xgemv: public Routine<T> {
  public:
-  Xgemv(CommandQueue &queue, Event &event);
+
+  // Members and methods from the base class
+  using Routine<T>::db_;
+  using Routine<T>::source_string_;
+  using Routine<T>::queue_;
+  using Routine<T>::GetProgramFromCache;
+  using Routine<T>::TestVectorX;
+  using Routine<T>::TestVectorY;
+  using Routine<T>::TestMatrixA;
+  using Routine<T>::RunKernel;
+  using Routine<T>::ErrorIn;
+
+  // Constructor
+  Xgemv(Queue &queue, Event &event);
 
   // 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 &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const Buffer &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 Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
                     const T beta,
-                    const Buffer &y_buffer, const size_t y_offset, const size_t y_inc);
+                    const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc);
 
  private:
   // Static variable to get the precision
diff --git a/include/internal/routines/level3/xgemm.h b/include/internal/routines/level3/xgemm.h
index 7ad4fcfb..a0c8b595 100644
--- a/include/internal/routines/level3/xgemm.h
+++ b/include/internal/routines/level3/xgemm.h
@@ -21,18 +21,33 @@ namespace clblast {
 
 // See comment at top of file for a description of the class
 template <typename T>
-class Xgemm: public Routine {
+class Xgemm: public Routine<T> {
  public:
-  Xgemm(CommandQueue &queue, Event &event);
+
+  // Members and methods from the base class
+  using Routine<T>::db_;
+  using Routine<T>::source_string_;
+  using Routine<T>::queue_;
+  using Routine<T>::context_;
+  using Routine<T>::GetProgramFromCache;
+  using Routine<T>::PadCopyTransposeMatrix;
+  using Routine<T>::TestMatrixA;
+  using Routine<T>::TestMatrixB;
+  using Routine<T>::TestMatrixC;
+  using Routine<T>::RunKernel;
+  using Routine<T>::ErrorIn;
+
+  // Constructor
+  Xgemm(Queue &queue, Event &event);
 
   // 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 &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const Buffer &b_buffer, const size_t b_offset, const size_t b_ld,
+                    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 &c_buffer, const size_t c_offset, const size_t c_ld);
+                    const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld);
 
  private:
   // Static variable to get the precision
diff --git a/include/internal/routines/level3/xhemm.h b/include/internal/routines/level3/xhemm.h
index 6cc9d9ec..5f1e8723 100644
--- a/include/internal/routines/level3/xhemm.h
+++ b/include/internal/routines/level3/xhemm.h
@@ -25,30 +25,28 @@ template <typename T>
 class Xhemm: public Xgemm<T> {
  public:
 
-  // Uses several variables from the Routine class
-  using Routine::db_;
-  using Routine::context_;
-
-  // Uses several helper functions from the Routine class
-  using Routine::RunKernel;
-  using Routine::ErrorIn;
-  using Routine::TestMatrixA;
-  using Routine::GetProgramFromCache;
+  // Members and methods from the base class
+  using Routine<T>::db_;
+  using Routine<T>::context_;
+  using Routine<T>::GetProgramFromCache;
+  using Routine<T>::TestMatrixA;
+  using Routine<T>::RunKernel;
+  using Routine<T>::ErrorIn;
 
   // Uses the regular Xgemm routine
   using Xgemm<T>::DoGemm;
 
   // Constructor
-  Xhemm(CommandQueue &queue, Event &event);
+  Xhemm(Queue &queue, Event &event);
 
   // 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 &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const Buffer &b_buffer, const size_t b_offset, const size_t b_ld,
+                    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 &c_buffer, const size_t c_offset, const size_t c_ld);
+                    const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld);
 };
 
 // =================================================================================================
diff --git a/include/internal/routines/level3/xher2k.h b/include/internal/routines/level3/xher2k.h
index 1836a812..9e961d23 100644
--- a/include/internal/routines/level3/xher2k.h
+++ b/include/internal/routines/level3/xher2k.h
@@ -23,18 +23,33 @@ namespace clblast {
 
 // See comment at top of file for a description of the class
 template <typename T, typename U>
-class Xher2k: public Routine {
+class Xher2k: public Routine<T> {
  public:
-  Xher2k(CommandQueue &queue, Event &event);
+
+  // Members and methods from the base class
+  using Routine<T>::db_;
+  using Routine<T>::source_string_;
+  using Routine<T>::queue_;
+  using Routine<T>::context_;
+  using Routine<T>::GetProgramFromCache;
+  using Routine<T>::PadCopyTransposeMatrix;
+  using Routine<T>::TestMatrixA;
+  using Routine<T>::TestMatrixB;
+  using Routine<T>::TestMatrixC;
+  using Routine<T>::RunKernel;
+  using Routine<T>::ErrorIn;
+
+  // Constructor
+  Xher2k(Queue &queue, Event &event);
 
   // 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 &a_buffer, const size_t a_offset, const size_t a_ld,
-                     const Buffer &b_buffer, const size_t b_offset, const size_t b_ld,
+                     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 &c_buffer, const size_t c_offset, const size_t c_ld);
+                     const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld);
 
  private:
   // Static variable to get the precision
diff --git a/include/internal/routines/level3/xherk.h b/include/internal/routines/level3/xherk.h
index 9b361254..f285a71c 100644
--- a/include/internal/routines/level3/xherk.h
+++ b/include/internal/routines/level3/xherk.h
@@ -23,17 +23,31 @@ namespace clblast {
 
 // See comment at top of file for a description of the class
 template <typename T, typename U>
-class Xherk: public Routine {
+class Xherk: public Routine<T> {
  public:
-  Xherk(CommandQueue &queue, Event &event);
+
+  // Members and methods from the base class
+  using Routine<T>::db_;
+  using Routine<T>::source_string_;
+  using Routine<T>::queue_;
+  using Routine<T>::context_;
+  using Routine<T>::GetProgramFromCache;
+  using Routine<T>::PadCopyTransposeMatrix;
+  using Routine<T>::TestMatrixA;
+  using Routine<T>::TestMatrixC;
+  using Routine<T>::RunKernel;
+  using Routine<T>::ErrorIn;
+
+  // Constructor
+  Xherk(Queue &queue, Event &event);
 
   // 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 &a_buffer, const size_t a_offset, const size_t a_ld,
+                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
                     const U beta,
-                    const Buffer &c_buffer, const size_t c_offset, const size_t c_ld);
+                    const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld);
 
  private:
   // Static variable to get the precision
diff --git a/include/internal/routines/level3/xsymm.h b/include/internal/routines/level3/xsymm.h
index 2028ceea..9ed3c722 100644
--- a/include/internal/routines/level3/xsymm.h
+++ b/include/internal/routines/level3/xsymm.h
@@ -27,30 +27,28 @@ template <typename T>
 class Xsymm: public Xgemm<T> {
  public:
 
-  // Uses several variables from the Routine class
-  using Routine::db_;
-  using Routine::context_;
-
-  // Uses several helper functions from the Routine class
-  using Routine::RunKernel;
-  using Routine::ErrorIn;
-  using Routine::TestMatrixA;
-  using Routine::GetProgramFromCache;
+  // Members and methods from the base class
+  using Routine<T>::db_;
+  using Routine<T>::context_;
+  using Routine<T>::GetProgramFromCache;
+  using Routine<T>::TestMatrixA;
+  using Routine<T>::RunKernel;
+  using Routine<T>::ErrorIn;
 
   // Uses the regular Xgemm routine
   using Xgemm<T>::DoGemm;
 
   // Constructor
-  Xsymm(CommandQueue &queue, Event &event);
+  Xsymm(Queue &queue, Event &event);
 
   // 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 &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const Buffer &b_buffer, const size_t b_offset, const size_t b_ld,
+                    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 &c_buffer, const size_t c_offset, const size_t c_ld);
+                    const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld);
 };
 
 // =================================================================================================
diff --git a/include/internal/routines/level3/xsyr2k.h b/include/internal/routines/level3/xsyr2k.h
index 6259313c..85936658 100644
--- a/include/internal/routines/level3/xsyr2k.h
+++ b/include/internal/routines/level3/xsyr2k.h
@@ -23,18 +23,33 @@ namespace clblast {
 
 // See comment at top of file for a description of the class
 template <typename T>
-class Xsyr2k: public Routine {
+class Xsyr2k: public Routine<T> {
  public:
-  Xsyr2k(CommandQueue &queue, Event &event);
+
+  // Members and methods from the base class
+  using Routine<T>::db_;
+  using Routine<T>::source_string_;
+  using Routine<T>::queue_;
+  using Routine<T>::context_;
+  using Routine<T>::GetProgramFromCache;
+  using Routine<T>::PadCopyTransposeMatrix;
+  using Routine<T>::TestMatrixA;
+  using Routine<T>::TestMatrixB;
+  using Routine<T>::TestMatrixC;
+  using Routine<T>::RunKernel;
+  using Routine<T>::ErrorIn;
+
+  // Constructor
+  Xsyr2k(Queue &queue, Event &event);
 
   // 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 &a_buffer, const size_t a_offset, const size_t a_ld,
-                     const Buffer &b_buffer, const size_t b_offset, const size_t b_ld,
+                     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 &c_buffer, const size_t c_offset, const size_t c_ld);
+                     const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld);
 
  private:
   // Static variable to get the precision
diff --git a/include/internal/routines/level3/xsyrk.h b/include/internal/routines/level3/xsyrk.h
index 3dab731f..14d51a58 100644
--- a/include/internal/routines/level3/xsyrk.h
+++ b/include/internal/routines/level3/xsyrk.h
@@ -25,17 +25,31 @@ namespace clblast {
 
 // See comment at top of file for a description of the class
 template <typename T>
-class Xsyrk: public Routine {
+class Xsyrk: public Routine<T> {
  public:
-  Xsyrk(CommandQueue &queue, Event &event);
+
+  // Members and methods from the base class
+  using Routine<T>::db_;
+  using Routine<T>::source_string_;
+  using Routine<T>::queue_;
+  using Routine<T>::context_;
+  using Routine<T>::GetProgramFromCache;
+  using Routine<T>::PadCopyTransposeMatrix;
+  using Routine<T>::TestMatrixA;
+  using Routine<T>::TestMatrixC;
+  using Routine<T>::RunKernel;
+  using Routine<T>::ErrorIn;
+
+  // Constructor
+  Xsyrk(Queue &queue, Event &event);
 
   // 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 &a_buffer, const size_t a_offset, const size_t a_ld,
+                    const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
                     const T beta,
-                    const Buffer &c_buffer, const size_t c_offset, const size_t c_ld);
+                    const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld);
 
  private:
   // Static variable to get the precision
diff --git a/include/internal/routines/level3/xtrmm.h b/include/internal/routines/level3/xtrmm.h
index 4f49bebd..d8ac60fd 100644
--- a/include/internal/routines/level3/xtrmm.h
+++ b/include/internal/routines/level3/xtrmm.h
@@ -26,29 +26,27 @@ template <typename T>
 class Xtrmm: public Xgemm<T> {
  public:
 
-  // Uses several variables from the Routine class
-  using Routine::db_;
-  using Routine::context_;
-
-  // Uses several helper functions from the Routine class
-  using Routine::RunKernel;
-  using Routine::ErrorIn;
-  using Routine::TestMatrixA;
-  using Routine::GetProgramFromCache;
+  // Members and methods from the base class
+  using Routine<T>::db_;
+  using Routine<T>::context_;
+  using Routine<T>::GetProgramFromCache;
+  using Routine<T>::TestMatrixA;
+  using Routine<T>::RunKernel;
+  using Routine<T>::ErrorIn;
 
   // Uses the regular Xgemm routine
   using Xgemm<T>::DoGemm;
 
   // Constructor
-  Xtrmm(CommandQueue &queue, Event &event);
+  Xtrmm(Queue &queue, Event &event);
 
   // 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 &a_buffer, const size_t a_offset, const size_t a_ld,
-                    const Buffer &b_buffer, const size_t b_offset, const size_t b_ld);
+                    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);
 };
 
 // =================================================================================================