From 2b56c2c60325f02bc695cbb968049cc09205c713 Mon Sep 17 00:00:00 2001 From: CNugteren Date: Sat, 26 Sep 2015 16:58:03 +0200 Subject: Added TRMV/TBMV/TPMV routines --- src/routines/level2/xtbmv.cc | 81 ++++++++++++++++++++++++++++++++++++++++++++ src/routines/level2/xtpmv.cc | 81 ++++++++++++++++++++++++++++++++++++++++++++ src/routines/level2/xtrmv.cc | 81 ++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 243 insertions(+) create mode 100644 src/routines/level2/xtbmv.cc create mode 100644 src/routines/level2/xtpmv.cc create mode 100644 src/routines/level2/xtrmv.cc (limited to 'src/routines') diff --git a/src/routines/level2/xtbmv.cc b/src/routines/level2/xtbmv.cc new file mode 100644 index 00000000..2e1aebff --- /dev/null +++ b/src/routines/level2/xtbmv.cc @@ -0,0 +1,81 @@ + +// ================================================================================================= +// 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 +// +// This file implements the Xtbmv class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/level2/xtbmv.h" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xtbmv::Xtbmv(Queue &queue, Event &event, const std::string &name): + Xgemv(queue, event, name) { +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xtbmv::DoTbmv(const Layout layout, const Triangle triangle, + const Transpose a_transpose, const Diagonal diagonal, + const size_t n, const size_t k, + 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) { + + // Creates a copy of X: a temporary scratch buffer + auto scratch_buffer = Buffer(context_, n*x_inc + x_offset); + try { + x_buffer.CopyTo(queue_, n*x_inc + x_offset, scratch_buffer); + } catch (...) { } // Continues: error-code is returned in MatVec + + // The data is either in the upper or lower triangle + size_t is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) || + (triangle == Triangle::kLower && layout == Layout::kRowMajor)); + + // Adds '2' to the parameter if the diagonal is unit + auto parameter = (diagonal == Diagonal::kUnit) ? is_upper + 2 : is_upper; + + // Runs the generic matrix-vector multiplication, disabling the use of fast vectorized kernels. + // The specific triangular banded matrix-accesses are implemented in the kernel guarded by the + // ROUTINE_TBMV define. + auto fast_kernels = false; + auto status = MatVec(layout, a_transpose, + n, n, static_cast(1), + a_buffer, a_offset, a_ld, + scratch_buffer, x_offset, x_inc, static_cast(0), + x_buffer, x_offset, x_inc, + fast_kernels, fast_kernels, + parameter, false, k, 0); + + // Returns the proper error code (renames vector Y to X) + switch(status) { + case StatusCode::kInvalidVectorY: return StatusCode::kInvalidVectorX; + case StatusCode::kInvalidIncrementY: return StatusCode::kInvalidIncrementX; + case StatusCode::kInsufficientMemoryY: return StatusCode::kInsufficientMemoryX; + default: return status; + } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xtbmv; +template class Xtbmv; +template class Xtbmv; +template class Xtbmv; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level2/xtpmv.cc b/src/routines/level2/xtpmv.cc new file mode 100644 index 00000000..aa0e099b --- /dev/null +++ b/src/routines/level2/xtpmv.cc @@ -0,0 +1,81 @@ + +// ================================================================================================= +// 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 +// +// This file implements the Xtpmv class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/level2/xtpmv.h" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xtpmv::Xtpmv(Queue &queue, Event &event, const std::string &name): + Xgemv(queue, event, name) { +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xtpmv::DoTpmv(const Layout layout, const Triangle triangle, + const Transpose a_transpose, const Diagonal diagonal, + const size_t n, + const Buffer &ap_buffer, const size_t ap_offset, + const Buffer &x_buffer, const size_t x_offset, const size_t x_inc) { + + // Creates a copy of X: a temporary scratch buffer + auto scratch_buffer = Buffer(context_, n*x_inc + x_offset); + try { + x_buffer.CopyTo(queue_, n*x_inc + x_offset, scratch_buffer); + } catch (...) { } // Continues: error-code is returned in MatVec + + // The data is either in the upper or lower triangle + size_t is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) || + (triangle == Triangle::kLower && layout == Layout::kRowMajor)); + + // Adds '2' to the parameter if the diagonal is unit + auto parameter = (diagonal == Diagonal::kUnit) ? is_upper + 2 : is_upper; + + // Runs the generic matrix-vector multiplication, disabling the use of fast vectorized kernels. + // The specific triangular packed matrix-accesses are implemented in the kernel guarded by the + // ROUTINE_TPMV define. + auto fast_kernels = false; + auto status = MatVec(layout, a_transpose, + n, n, static_cast(1), + ap_buffer, ap_offset, n, + scratch_buffer, x_offset, x_inc, static_cast(0), + x_buffer, x_offset, x_inc, + fast_kernels, fast_kernels, + parameter, true, 0, 0); + + // Returns the proper error code (renames vector Y to X) + switch(status) { + case StatusCode::kInvalidVectorY: return StatusCode::kInvalidVectorX; + case StatusCode::kInvalidIncrementY: return StatusCode::kInvalidIncrementX; + case StatusCode::kInsufficientMemoryY: return StatusCode::kInsufficientMemoryX; + default: return status; + } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xtpmv; +template class Xtpmv; +template class Xtpmv; +template class Xtpmv; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level2/xtrmv.cc b/src/routines/level2/xtrmv.cc new file mode 100644 index 00000000..94424743 --- /dev/null +++ b/src/routines/level2/xtrmv.cc @@ -0,0 +1,81 @@ + +// ================================================================================================= +// 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 +// +// This file implements the Xtrmv class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/level2/xtrmv.h" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xtrmv::Xtrmv(Queue &queue, Event &event, const std::string &name): + Xgemv(queue, event, name) { +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xtrmv::DoTrmv(const Layout layout, const Triangle triangle, + const Transpose a_transpose, const Diagonal diagonal, + const size_t n, + 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) { + + // Creates a copy of X: a temporary scratch buffer + auto scratch_buffer = Buffer(context_, n*x_inc + x_offset); + try { + x_buffer.CopyTo(queue_, n*x_inc + x_offset, scratch_buffer); + } catch (...) { } // Continues: error-code is returned in MatVec + + // The data is either in the upper or lower triangle + size_t is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) || + (triangle == Triangle::kLower && layout == Layout::kRowMajor)); + + // Adds '2' to the parameter if the diagonal is unit + auto parameter = (diagonal == Diagonal::kUnit) ? is_upper + 2 : is_upper; + + // Runs the generic matrix-vector multiplication, disabling the use of fast vectorized kernels. + // The specific triangular matrix-accesses are implemented in the kernel guarded by the + // ROUTINE_TRMV define. + auto fast_kernels = false; + auto status = MatVec(layout, a_transpose, + n, n, static_cast(1), + a_buffer, a_offset, a_ld, + scratch_buffer, x_offset, x_inc, static_cast(0), + x_buffer, x_offset, x_inc, + fast_kernels, fast_kernels, + parameter, false, 0, 0); + + // Returns the proper error code (renames vector Y to X) + switch(status) { + case StatusCode::kInvalidVectorY: return StatusCode::kInvalidVectorX; + case StatusCode::kInvalidIncrementY: return StatusCode::kInvalidIncrementX; + case StatusCode::kInsufficientMemoryY: return StatusCode::kInsufficientMemoryX; + default: return status; + } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xtrmv; +template class Xtrmv; +template class Xtrmv; +template class Xtrmv; + +// ================================================================================================= +} // namespace clblast -- cgit v1.2.3