From ff0c54c3865b45eff807315262e73d3f01cb19c3 Mon Sep 17 00:00:00 2001 From: CNugteren Date: Sat, 22 Aug 2015 17:11:20 +0200 Subject: Added the XSWAP, XSCAL and XCOPY level-1 routines --- src/routines/level1/xcopy.cc | 117 +++++++++++++++++++++++++++++++++++++++++++ src/routines/level1/xscal.cc | 111 ++++++++++++++++++++++++++++++++++++++++ src/routines/level1/xswap.cc | 117 +++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 345 insertions(+) create mode 100644 src/routines/level1/xcopy.cc create mode 100644 src/routines/level1/xscal.cc create mode 100644 src/routines/level1/xswap.cc (limited to 'src/routines/level1') diff --git a/src/routines/level1/xcopy.cc b/src/routines/level1/xcopy.cc new file mode 100644 index 00000000..52e029b9 --- /dev/null +++ b/src/routines/level1/xcopy.cc @@ -0,0 +1,117 @@ + +// ================================================================================================= +// 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 Xcopy class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/level1/xcopy.h" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Specific implementations to get the memory-type based on a template argument +template <> const Precision Xcopy::precision_ = Precision::kSingle; +template <> const Precision Xcopy::precision_ = Precision::kDouble; +template <> const Precision Xcopy::precision_ = Precision::kComplexSingle; +template <> const Precision Xcopy::precision_ = Precision::kComplexDouble; + +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xcopy::Xcopy(Queue &queue, Event &event): + Routine(queue, event, "COPY", {"Xaxpy"}, precision_) { + source_string_ = + #include "../../kernels/level1/level1.opencl" + #include "../../kernels/level1/xcopy.opencl" + ; +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xcopy::DoCopy(const size_t n, + 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) { + + // Makes sure all dimensions are larger than zero + if (n == 0) { return StatusCode::kInvalidDimension; } + + // Tests the vectors for validity + auto status = TestVectorX(n, x_buffer, x_offset, x_inc, sizeof(T)); + if (ErrorIn(status)) { return status; } + status = TestVectorY(n, y_buffer, y_offset, y_inc, sizeof(T)); + if (ErrorIn(status)) { return status; } + + // Determines whether or not the fast-version can be used + bool use_fast_kernel = (x_offset == 0) && (x_inc == 1) && + (y_offset == 0) && (y_inc == 1) && + IsMultiple(n, db_["WGS"]*db_["WPT"]*db_["VW"]); + + // If possible, run the fast-version of the kernel + auto kernel_name = (use_fast_kernel) ? "XcopyFast" : "Xcopy"; + + // Retrieves the Xcopy kernel from the compiled binary + try { + auto& program = GetProgramFromCache(); + auto kernel = Kernel(program, kernel_name); + + // Sets the kernel arguments + if (use_fast_kernel) { + kernel.SetArgument(0, static_cast(n)); + kernel.SetArgument(1, x_buffer()); + kernel.SetArgument(2, y_buffer()); + } + else { + kernel.SetArgument(0, static_cast(n)); + kernel.SetArgument(1, x_buffer()); + kernel.SetArgument(2, static_cast(x_offset)); + kernel.SetArgument(3, static_cast(x_inc)); + kernel.SetArgument(4, y_buffer()); + kernel.SetArgument(5, static_cast(y_offset)); + kernel.SetArgument(6, static_cast(y_inc)); + } + + // Launches the kernel + if (use_fast_kernel) { + auto global = std::vector{CeilDiv(n, db_["WPT"]*db_["VW"])}; + auto local = std::vector{db_["WGS"]}; + status = RunKernel(kernel, global, local); + } + else { + auto n_ceiled = Ceil(n, db_["WGS"]*db_["WPT"]); + auto global = std::vector{n_ceiled/db_["WPT"]}; + auto local = std::vector{db_["WGS"]}; + status = RunKernel(kernel, global, local); + } + if (ErrorIn(status)) { return status; } + + // Waits for all kernels to finish + queue_.Finish(); + + // Succesfully finished the computation + return StatusCode::kSuccess; + } catch (...) { return StatusCode::kInvalidKernel; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xcopy; +template class Xcopy; +template class Xcopy; +template class Xcopy; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level1/xscal.cc b/src/routines/level1/xscal.cc new file mode 100644 index 00000000..13e1080c --- /dev/null +++ b/src/routines/level1/xscal.cc @@ -0,0 +1,111 @@ + +// ================================================================================================= +// 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 Xscal class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/level1/xscal.h" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Specific implementations to get the memory-type based on a template argument +template <> const Precision Xscal::precision_ = Precision::kSingle; +template <> const Precision Xscal::precision_ = Precision::kDouble; +template <> const Precision Xscal::precision_ = Precision::kComplexSingle; +template <> const Precision Xscal::precision_ = Precision::kComplexDouble; + +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xscal::Xscal(Queue &queue, Event &event): + Routine(queue, event, "SCAL", {"Xaxpy"}, precision_) { + source_string_ = + #include "../../kernels/level1/level1.opencl" + #include "../../kernels/level1/xscal.opencl" + ; +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xscal::DoScal(const size_t n, const T alpha, + const Buffer &x_buffer, const size_t x_offset, const size_t x_inc) { + + // Makes sure all dimensions are larger than zero + if (n == 0) { return StatusCode::kInvalidDimension; } + + // Tests the vector for validity + auto status = TestVectorX(n, x_buffer, x_offset, x_inc, sizeof(T)); + if (ErrorIn(status)) { return status; } + + // Determines whether or not the fast-version can be used + bool use_fast_kernel = (x_offset == 0) && (x_inc == 1) && + IsMultiple(n, db_["WGS"]*db_["WPT"]*db_["VW"]); + + // If possible, run the fast-version of the kernel + auto kernel_name = (use_fast_kernel) ? "XscalFast" : "Xscal"; + + // Retrieves the Xscal kernel from the compiled binary + try { + auto& program = GetProgramFromCache(); + auto kernel = Kernel(program, kernel_name); + + // Sets the kernel arguments + if (use_fast_kernel) { + kernel.SetArgument(0, static_cast(n)); + kernel.SetArgument(1, alpha); + kernel.SetArgument(2, x_buffer()); + } + else { + kernel.SetArgument(0, static_cast(n)); + kernel.SetArgument(1, alpha); + kernel.SetArgument(2, x_buffer()); + kernel.SetArgument(3, static_cast(x_offset)); + kernel.SetArgument(4, static_cast(x_inc)); + } + + // Launches the kernel + if (use_fast_kernel) { + auto global = std::vector{CeilDiv(n, db_["WPT"]*db_["VW"])}; + auto local = std::vector{db_["WGS"]}; + status = RunKernel(kernel, global, local); + } + else { + auto n_ceiled = Ceil(n, db_["WGS"]*db_["WPT"]); + auto global = std::vector{n_ceiled/db_["WPT"]}; + auto local = std::vector{db_["WGS"]}; + status = RunKernel(kernel, global, local); + } + if (ErrorIn(status)) { return status; } + + // Waits for all kernels to finish + queue_.Finish(); + + // Succesfully finished the computation + return StatusCode::kSuccess; + } catch (...) { return StatusCode::kInvalidKernel; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xscal; +template class Xscal; +template class Xscal; +template class Xscal; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level1/xswap.cc b/src/routines/level1/xswap.cc new file mode 100644 index 00000000..b22b3bdb --- /dev/null +++ b/src/routines/level1/xswap.cc @@ -0,0 +1,117 @@ + +// ================================================================================================= +// 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 Xswap class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/level1/xswap.h" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Specific implementations to get the memory-type based on a template argument +template <> const Precision Xswap::precision_ = Precision::kSingle; +template <> const Precision Xswap::precision_ = Precision::kDouble; +template <> const Precision Xswap::precision_ = Precision::kComplexSingle; +template <> const Precision Xswap::precision_ = Precision::kComplexDouble; + +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xswap::Xswap(Queue &queue, Event &event): + Routine(queue, event, "SWAP", {"Xaxpy"}, precision_) { + source_string_ = + #include "../../kernels/level1/level1.opencl" + #include "../../kernels/level1/xswap.opencl" + ; +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xswap::DoSwap(const size_t n, + 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) { + + // Makes sure all dimensions are larger than zero + if (n == 0) { return StatusCode::kInvalidDimension; } + + // Tests the vectors for validity + auto status = TestVectorX(n, x_buffer, x_offset, x_inc, sizeof(T)); + if (ErrorIn(status)) { return status; } + status = TestVectorY(n, y_buffer, y_offset, y_inc, sizeof(T)); + if (ErrorIn(status)) { return status; } + + // Determines whether or not the fast-version can be used + bool use_fast_kernel = (x_offset == 0) && (x_inc == 1) && + (y_offset == 0) && (y_inc == 1) && + IsMultiple(n, db_["WGS"]*db_["WPT"]*db_["VW"]); + + // If possible, run the fast-version of the kernel + auto kernel_name = (use_fast_kernel) ? "XswapFast" : "Xswap"; + + // Retrieves the Xswap kernel from the compiled binary + try { + auto& program = GetProgramFromCache(); + auto kernel = Kernel(program, kernel_name); + + // Sets the kernel arguments + if (use_fast_kernel) { + kernel.SetArgument(0, static_cast(n)); + kernel.SetArgument(1, x_buffer()); + kernel.SetArgument(2, y_buffer()); + } + else { + kernel.SetArgument(0, static_cast(n)); + kernel.SetArgument(1, x_buffer()); + kernel.SetArgument(2, static_cast(x_offset)); + kernel.SetArgument(3, static_cast(x_inc)); + kernel.SetArgument(4, y_buffer()); + kernel.SetArgument(5, static_cast(y_offset)); + kernel.SetArgument(6, static_cast(y_inc)); + } + + // Launches the kernel + if (use_fast_kernel) { + auto global = std::vector{CeilDiv(n, db_["WPT"]*db_["VW"])}; + auto local = std::vector{db_["WGS"]}; + status = RunKernel(kernel, global, local); + } + else { + auto n_ceiled = Ceil(n, db_["WGS"]*db_["WPT"]); + auto global = std::vector{n_ceiled/db_["WPT"]}; + auto local = std::vector{db_["WGS"]}; + status = RunKernel(kernel, global, local); + } + if (ErrorIn(status)) { return status; } + + // Waits for all kernels to finish + queue_.Finish(); + + // Succesfully finished the computation + return StatusCode::kSuccess; + } catch (...) { return StatusCode::kInvalidKernel; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xswap; +template class Xswap; +template class Xswap; +template class Xswap; + +// ================================================================================================= +} // namespace clblast -- cgit v1.2.3