diff options
Diffstat (limited to 'src/routines/level2')
44 files changed, 1066 insertions, 22 deletions
diff --git a/src/routines/level2/xgbmv.cc b/src/routines/level2/xgbmv.cc index 7a30c34a..ea4f001c 100644 --- a/src/routines/level2/xgbmv.cc +++ b/src/routines/level2/xgbmv.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xgbmv.h" +#include "routines/level2/xgbmv.hpp" #include <string> #include <vector> diff --git a/src/routines/level2/xgbmv.hpp b/src/routines/level2/xgbmv.hpp new file mode 100644 index 00000000..686ab642 --- /dev/null +++ b/src/routines/level2/xgbmv.hpp @@ -0,0 +1,49 @@ + +// ================================================================================================= +// 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 "routines/level2/xgemv.hpp" + +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/src/routines/level2/xgemv.cc b/src/routines/level2/xgemv.cc index ccadd131..21fb397c 100644 --- a/src/routines/level2/xgemv.cc +++ b/src/routines/level2/xgemv.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xgemv.h" +#include "routines/level2/xgemv.hpp" #include <string> #include <vector> diff --git a/src/routines/level2/xgemv.hpp b/src/routines/level2/xgemv.hpp new file mode 100644 index 00000000..e9afec8d --- /dev/null +++ b/src/routines/level2/xgemv.hpp @@ -0,0 +1,56 @@ + +// ================================================================================================= +// 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 "routine.hpp" + +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/src/routines/level2/xger.cc b/src/routines/level2/xger.cc index 6ceaa00e..353047d2 100644 --- a/src/routines/level2/xger.cc +++ b/src/routines/level2/xger.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xger.h" +#include "routines/level2/xger.hpp" #include <string> #include <vector> diff --git a/src/routines/level2/xger.hpp b/src/routines/level2/xger.hpp new file mode 100644 index 00000000..3c6abe44 --- /dev/null +++ b/src/routines/level2/xger.hpp @@ -0,0 +1,43 @@ + +// ================================================================================================= +// 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 "routine.hpp" + +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/src/routines/level2/xgerc.cc b/src/routines/level2/xgerc.cc index 73284b52..d9feda97 100644 --- a/src/routines/level2/xgerc.cc +++ b/src/routines/level2/xgerc.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xgerc.h" +#include "routines/level2/xgerc.hpp" #include <string> diff --git a/src/routines/level2/xgerc.hpp b/src/routines/level2/xgerc.hpp new file mode 100644 index 00000000..f1d04dfd --- /dev/null +++ b/src/routines/level2/xgerc.hpp @@ -0,0 +1,46 @@ + +// ================================================================================================= +// 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 "routines/level2/xger.hpp" + +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/src/routines/level2/xgeru.cc b/src/routines/level2/xgeru.cc index 7730d6a5..da9e91c2 100644 --- a/src/routines/level2/xgeru.cc +++ b/src/routines/level2/xgeru.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xgeru.h" +#include "routines/level2/xgeru.hpp" #include <string> diff --git a/src/routines/level2/xgeru.hpp b/src/routines/level2/xgeru.hpp new file mode 100644 index 00000000..fb50e917 --- /dev/null +++ b/src/routines/level2/xgeru.hpp @@ -0,0 +1,46 @@ + +// ================================================================================================= +// 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 "routines/level2/xger.hpp" + +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/src/routines/level2/xhbmv.cc b/src/routines/level2/xhbmv.cc index 58591b50..f6c0e3c4 100644 --- a/src/routines/level2/xhbmv.cc +++ b/src/routines/level2/xhbmv.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xhbmv.h" +#include "routines/level2/xhbmv.hpp" #include <string> #include <vector> diff --git a/src/routines/level2/xhbmv.hpp b/src/routines/level2/xhbmv.hpp new file mode 100644 index 00000000..d668eb88 --- /dev/null +++ b/src/routines/level2/xhbmv.hpp @@ -0,0 +1,49 @@ + +// ================================================================================================= +// 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 "routines/level2/xgemv.hpp" + +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/src/routines/level2/xhemv.cc b/src/routines/level2/xhemv.cc index b4ef0fa4..2cbcf7b4 100644 --- a/src/routines/level2/xhemv.cc +++ b/src/routines/level2/xhemv.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xhemv.h" +#include "routines/level2/xhemv.hpp" #include <string> #include <vector> diff --git a/src/routines/level2/xhemv.hpp b/src/routines/level2/xhemv.hpp new file mode 100644 index 00000000..8e062fd3 --- /dev/null +++ b/src/routines/level2/xhemv.hpp @@ -0,0 +1,49 @@ + +// ================================================================================================= +// 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 "routines/level2/xgemv.hpp" + +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/src/routines/level2/xher.cc b/src/routines/level2/xher.cc index 939e17bb..ed8ba9e9 100644 --- a/src/routines/level2/xher.cc +++ b/src/routines/level2/xher.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xher.h" +#include "routines/level2/xher.hpp" #include <string> diff --git a/src/routines/level2/xher.hpp b/src/routines/level2/xher.hpp new file mode 100644 index 00000000..9ff6bf3f --- /dev/null +++ b/src/routines/level2/xher.hpp @@ -0,0 +1,46 @@ + +// ================================================================================================= +// 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 "routine.hpp" + +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/src/routines/level2/xher2.cc b/src/routines/level2/xher2.cc index 95dbd87a..50572cea 100644 --- a/src/routines/level2/xher2.cc +++ b/src/routines/level2/xher2.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xher2.h" +#include "routines/level2/xher2.hpp" #include <string> diff --git a/src/routines/level2/xher2.hpp b/src/routines/level2/xher2.hpp new file mode 100644 index 00000000..8c53c047 --- /dev/null +++ b/src/routines/level2/xher2.hpp @@ -0,0 +1,44 @@ + +// ================================================================================================= +// 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 "routine.hpp" + +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/src/routines/level2/xhpmv.cc b/src/routines/level2/xhpmv.cc index 92686dbe..e6f82b34 100644 --- a/src/routines/level2/xhpmv.cc +++ b/src/routines/level2/xhpmv.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xhpmv.h" +#include "routines/level2/xhpmv.hpp" #include <string> #include <vector> diff --git a/src/routines/level2/xhpmv.hpp b/src/routines/level2/xhpmv.hpp new file mode 100644 index 00000000..b11192f9 --- /dev/null +++ b/src/routines/level2/xhpmv.hpp @@ -0,0 +1,49 @@ + +// ================================================================================================= +// 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 "routines/level2/xgemv.hpp" + +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/src/routines/level2/xhpr.cc b/src/routines/level2/xhpr.cc index 4b31ad09..225ebfe5 100644 --- a/src/routines/level2/xhpr.cc +++ b/src/routines/level2/xhpr.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xhpr.h" +#include "routines/level2/xhpr.hpp" #include <string> diff --git a/src/routines/level2/xhpr.hpp b/src/routines/level2/xhpr.hpp new file mode 100644 index 00000000..37801c68 --- /dev/null +++ b/src/routines/level2/xhpr.hpp @@ -0,0 +1,45 @@ + +// ================================================================================================= +// 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 "routines/level2/xher.hpp" + +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/src/routines/level2/xhpr2.cc b/src/routines/level2/xhpr2.cc index 9be24f43..85f9d3f9 100644 --- a/src/routines/level2/xhpr2.cc +++ b/src/routines/level2/xhpr2.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xhpr2.h" +#include "routines/level2/xhpr2.hpp" #include <string> diff --git a/src/routines/level2/xhpr2.hpp b/src/routines/level2/xhpr2.hpp new file mode 100644 index 00000000..d66dce55 --- /dev/null +++ b/src/routines/level2/xhpr2.hpp @@ -0,0 +1,46 @@ + +// ================================================================================================= +// 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 "routines/level2/xher2.hpp" + +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/src/routines/level2/xsbmv.cc b/src/routines/level2/xsbmv.cc index 66ba74e8..28730899 100644 --- a/src/routines/level2/xsbmv.cc +++ b/src/routines/level2/xsbmv.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xsbmv.h" +#include "routines/level2/xsbmv.hpp" #include <string> #include <vector> diff --git a/src/routines/level2/xsbmv.hpp b/src/routines/level2/xsbmv.hpp new file mode 100644 index 00000000..16c5e9a8 --- /dev/null +++ b/src/routines/level2/xsbmv.hpp @@ -0,0 +1,49 @@ + +// ================================================================================================= +// 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 "routines/level2/xgemv.hpp" + +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/src/routines/level2/xspmv.cc b/src/routines/level2/xspmv.cc index 589a97d4..f6651012 100644 --- a/src/routines/level2/xspmv.cc +++ b/src/routines/level2/xspmv.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xspmv.h" +#include "routines/level2/xspmv.hpp" #include <string> #include <vector> diff --git a/src/routines/level2/xspmv.hpp b/src/routines/level2/xspmv.hpp new file mode 100644 index 00000000..a0c69b85 --- /dev/null +++ b/src/routines/level2/xspmv.hpp @@ -0,0 +1,49 @@ + +// ================================================================================================= +// 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 "routines/level2/xgemv.hpp" + +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/src/routines/level2/xspr.cc b/src/routines/level2/xspr.cc index c556b920..a75fe9c3 100644 --- a/src/routines/level2/xspr.cc +++ b/src/routines/level2/xspr.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xspr.h" +#include "routines/level2/xspr.hpp" #include <string> diff --git a/src/routines/level2/xspr.hpp b/src/routines/level2/xspr.hpp new file mode 100644 index 00000000..6468c736 --- /dev/null +++ b/src/routines/level2/xspr.hpp @@ -0,0 +1,45 @@ + +// ================================================================================================= +// 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 "routines/level2/xher.hpp" + +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/src/routines/level2/xspr2.cc b/src/routines/level2/xspr2.cc index c4ad5dc4..c39a2eb4 100644 --- a/src/routines/level2/xspr2.cc +++ b/src/routines/level2/xspr2.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xspr2.h" +#include "routines/level2/xspr2.hpp" #include <string> diff --git a/src/routines/level2/xspr2.hpp b/src/routines/level2/xspr2.hpp new file mode 100644 index 00000000..693c56a1 --- /dev/null +++ b/src/routines/level2/xspr2.hpp @@ -0,0 +1,46 @@ + +// ================================================================================================= +// 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 "routines/level2/xher2.hpp" + +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/src/routines/level2/xsymv.cc b/src/routines/level2/xsymv.cc index 2a404a8a..648d2a3e 100644 --- a/src/routines/level2/xsymv.cc +++ b/src/routines/level2/xsymv.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xsymv.h" +#include "routines/level2/xsymv.hpp" #include <string> #include <vector> diff --git a/src/routines/level2/xsymv.hpp b/src/routines/level2/xsymv.hpp new file mode 100644 index 00000000..67815f2f --- /dev/null +++ b/src/routines/level2/xsymv.hpp @@ -0,0 +1,49 @@ + +// ================================================================================================= +// 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 "routines/level2/xgemv.hpp" + +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/src/routines/level2/xsyr.cc b/src/routines/level2/xsyr.cc index 892517d7..758d8f8f 100644 --- a/src/routines/level2/xsyr.cc +++ b/src/routines/level2/xsyr.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xsyr.h" +#include "routines/level2/xsyr.hpp" #include <string> diff --git a/src/routines/level2/xsyr.hpp b/src/routines/level2/xsyr.hpp new file mode 100644 index 00000000..20393454 --- /dev/null +++ b/src/routines/level2/xsyr.hpp @@ -0,0 +1,45 @@ + +// ================================================================================================= +// 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 "routines/level2/xher.hpp" + +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/src/routines/level2/xsyr2.cc b/src/routines/level2/xsyr2.cc index e6dfd158..6f43b219 100644 --- a/src/routines/level2/xsyr2.cc +++ b/src/routines/level2/xsyr2.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xsyr2.h" +#include "routines/level2/xsyr2.hpp" #include <string> diff --git a/src/routines/level2/xsyr2.hpp b/src/routines/level2/xsyr2.hpp new file mode 100644 index 00000000..1a8dcbe8 --- /dev/null +++ b/src/routines/level2/xsyr2.hpp @@ -0,0 +1,46 @@ + +// ================================================================================================= +// 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 "routines/level2/xher2.hpp" + +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/src/routines/level2/xtbmv.cc b/src/routines/level2/xtbmv.cc index 86e28dfb..e315c544 100644 --- a/src/routines/level2/xtbmv.cc +++ b/src/routines/level2/xtbmv.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xtbmv.h" +#include "routines/level2/xtbmv.hpp" #include <string> #include <vector> diff --git a/src/routines/level2/xtbmv.hpp b/src/routines/level2/xtbmv.hpp new file mode 100644 index 00000000..389e9705 --- /dev/null +++ b/src/routines/level2/xtbmv.hpp @@ -0,0 +1,49 @@ + +// ================================================================================================= +// 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 "routines/level2/xgemv.hpp" + +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/src/routines/level2/xtpmv.cc b/src/routines/level2/xtpmv.cc index 72445547..46811089 100644 --- a/src/routines/level2/xtpmv.cc +++ b/src/routines/level2/xtpmv.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xtpmv.h" +#include "routines/level2/xtpmv.hpp" #include <string> #include <vector> diff --git a/src/routines/level2/xtpmv.hpp b/src/routines/level2/xtpmv.hpp new file mode 100644 index 00000000..0e8cf1d2 --- /dev/null +++ b/src/routines/level2/xtpmv.hpp @@ -0,0 +1,49 @@ + +// ================================================================================================= +// 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 "routines/level2/xgemv.hpp" + +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/src/routines/level2/xtrmv.cc b/src/routines/level2/xtrmv.cc index df6f85a3..d2f24252 100644 --- a/src/routines/level2/xtrmv.cc +++ b/src/routines/level2/xtrmv.cc @@ -11,7 +11,7 @@ // // ================================================================================================= -#include "internal/routines/level2/xtrmv.h" +#include "routines/level2/xtrmv.hpp" #include <string> #include <vector> diff --git a/src/routines/level2/xtrmv.hpp b/src/routines/level2/xtrmv.hpp new file mode 100644 index 00000000..07dd7841 --- /dev/null +++ b/src/routines/level2/xtrmv.hpp @@ -0,0 +1,49 @@ + +// ================================================================================================= +// 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 "routines/level2/xgemv.hpp" + +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 |