From 61203453aaca4e47c05c598a673150522160ca87 Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Sun, 19 Jun 2016 13:55:49 +0200 Subject: Renamed all C++ source files to .cpp to match the .hpp extension better --- src/routines/level3/xgemm.cc | 223 -------------------------------------- src/routines/level3/xgemm.cpp | 223 ++++++++++++++++++++++++++++++++++++++ src/routines/level3/xhemm.cc | 134 ----------------------- src/routines/level3/xhemm.cpp | 134 +++++++++++++++++++++++ src/routines/level3/xher2k.cc | 241 ----------------------------------------- src/routines/level3/xher2k.cpp | 241 +++++++++++++++++++++++++++++++++++++++++ src/routines/level3/xherk.cc | 197 --------------------------------- src/routines/level3/xherk.cpp | 197 +++++++++++++++++++++++++++++++++ src/routines/level3/xsymm.cc | 137 ----------------------- src/routines/level3/xsymm.cpp | 137 +++++++++++++++++++++++ src/routines/level3/xsyr2k.cc | 210 ----------------------------------- src/routines/level3/xsyr2k.cpp | 210 +++++++++++++++++++++++++++++++++++ src/routines/level3/xsyrk.cc | 181 ------------------------------- src/routines/level3/xsyrk.cpp | 181 +++++++++++++++++++++++++++++++ src/routines/level3/xtrmm.cc | 140 ------------------------ src/routines/level3/xtrmm.cpp | 140 ++++++++++++++++++++++++ 16 files changed, 1463 insertions(+), 1463 deletions(-) delete mode 100644 src/routines/level3/xgemm.cc create mode 100644 src/routines/level3/xgemm.cpp delete mode 100644 src/routines/level3/xhemm.cc create mode 100644 src/routines/level3/xhemm.cpp delete mode 100644 src/routines/level3/xher2k.cc create mode 100644 src/routines/level3/xher2k.cpp delete mode 100644 src/routines/level3/xherk.cc create mode 100644 src/routines/level3/xherk.cpp delete mode 100644 src/routines/level3/xsymm.cc create mode 100644 src/routines/level3/xsymm.cpp delete mode 100644 src/routines/level3/xsyr2k.cc create mode 100644 src/routines/level3/xsyr2k.cpp delete mode 100644 src/routines/level3/xsyrk.cc create mode 100644 src/routines/level3/xsyrk.cpp delete mode 100644 src/routines/level3/xtrmm.cc create mode 100644 src/routines/level3/xtrmm.cpp (limited to 'src/routines/level3') diff --git a/src/routines/level3/xgemm.cc b/src/routines/level3/xgemm.cc deleted file mode 100644 index 9ea5559c..00000000 --- a/src/routines/level3/xgemm.cc +++ /dev/null @@ -1,223 +0,0 @@ - -// ================================================================================================= -// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This -// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max- -// width of 100 characters per line. -// -// Author(s): -// Cedric Nugteren -// -// This file implements the Xgemm class (see the header for information about the class). -// -// ================================================================================================= - -#include "routines/level3/xgemm.hpp" - -#include -#include - -namespace clblast { -// ================================================================================================= - -// Constructor: forwards to base class constructor -template -Xgemm::Xgemm(Queue &queue, EventPointer event, const std::string &name): - Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue()) { - source_string_ = - #include "../../kernels/level3/level3.opencl" - #include "../../kernels/level3/copy_fast.opencl" - #include "../../kernels/level3/copy_pad.opencl" - #include "../../kernels/level3/transpose_fast.opencl" - #include "../../kernels/level3/transpose_pad.opencl" - #include "../../kernels/level3/convert_symmetric.opencl" - #include "../../kernels/level3/convert_triangular.opencl" - #include "../../kernels/level3/convert_hermitian.opencl" - #include "../../kernels/level3/xgemm_part1.opencl" - #include "../../kernels/level3/xgemm_part2.opencl" - ; -} - -// ================================================================================================= - -// The main routine -template -StatusCode Xgemm::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 T beta, - const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { - - // Makes sure all dimensions are larger than zero - if ((m == 0) || (n == 0) || (k == 0)) { return StatusCode::kInvalidDimension; } - - // Computes whether or not the matrices are transposed in memory. This is based on their layout - // (row or column-major) and whether or not they are requested to be pre-transposed. Note - // that the Xgemm kernel expects either matrices A and C (in case of row-major) or B (in case of - // col-major) to be transformed, so transposing requirements are not the same as whether or not - // the matrix is actually transposed in memory. - const auto a_rotated = (layout == Layout::kColMajor && a_transpose != Transpose::kNo) || - (layout == Layout::kRowMajor && a_transpose == Transpose::kNo); - const auto b_rotated = (layout == Layout::kColMajor && b_transpose != Transpose::kNo) || - (layout == Layout::kRowMajor && b_transpose == Transpose::kNo); - const auto c_rotated = (layout == Layout::kRowMajor); - const auto a_do_transpose = a_rotated; - const auto b_do_transpose = !b_rotated; - const auto c_do_transpose = c_rotated; - - // In case of complex data-types, the transpose can also become a conjugate transpose - const auto a_conjugate = (a_transpose == Transpose::kConjugate); - const auto b_conjugate = (b_transpose == Transpose::kConjugate); - - // Computes the first and second dimensions of the 3 matrices taking into account whether the - // matrices are rotated or not - const auto a_one = (a_rotated) ? k : m; - const auto a_two = (a_rotated) ? m : k; - const auto b_one = (b_rotated) ? n : k; - const auto b_two = (b_rotated) ? k : n; - const auto c_one = (c_rotated) ? n : m; - const auto c_two = (c_rotated) ? m : n; - - // Tests three matrices (A, B, C) for validity, first from a perspective of the OpenCL buffers and - // their sizes, and then from a perspective of parameter values (e.g. m, n, k). Tests whether the - // OpenCL buffers are valid and non-zero and whether the OpenCL buffers have sufficient storage - // space. Also tests that the leading dimensions of: - // matrix A cannot be less than K when rotated, or less than M when not-rotated - // matrix B cannot be less than N when rotated, or less than K when not-rotated - // matrix C cannot be less than N when rotated, or less than M when not-rotated - auto status = TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld); - if (ErrorIn(status)) { return status; } - status = TestMatrixB(b_one, b_two, b_buffer, b_offset, b_ld); - if (ErrorIn(status)) { return status; } - status = TestMatrixC(c_one, c_two, c_buffer, c_offset, c_ld); - if (ErrorIn(status)) { return status; } - - // Calculates the ceiled versions of m, n, and k - const auto m_ceiled = Ceil(m, db_["MWG"]); - const auto n_ceiled = Ceil(n, db_["NWG"]); - const auto k_ceiled = Ceil(k, db_["KWG"]); - - // The padded/transposed input/output matrices: if memory allocation fails, throw an exception - try { - - // Loads the program from the database - const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); - - // Determines whether or not temporary matrices are needed - auto a_no_temp = a_one == m_ceiled && a_two == k_ceiled && a_ld == m_ceiled && a_offset == 0 && - a_do_transpose == false && a_conjugate == false; - auto b_no_temp = b_one == n_ceiled && b_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && - b_do_transpose == false && b_conjugate == false; - auto c_no_temp = c_one == m_ceiled && c_two == n_ceiled && c_ld == m_ceiled && c_offset == 0 && - c_do_transpose == false; - - // Creates the temporary matrices - const auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, k_ceiled*m_ceiled); - const auto b_temp = (b_no_temp) ? b_buffer : Buffer(context_, k_ceiled*n_ceiled); - const auto c_temp = (c_no_temp) ? c_buffer : Buffer(context_, m_ceiled*n_ceiled); - - // Upload the scalar arguments as constant buffers to the device (needed for half-precision) - auto alpha_buffer = Buffer(context_, 1); - auto beta_buffer = Buffer(context_, 1); - alpha_buffer.Write(queue_, 1, &alpha); - beta_buffer.Write(queue_, 1, &beta); - - // Events of all kernels (including pre/post processing kernels) - auto eventWaitList = std::vector(); - auto emptyEventList = std::vector(); - - // Runs the pre-processing kernel for matrix A. This transposes the matrix, but also pads zeros - // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In - // case nothing has to be done, these kernels can be skipped. - if (!a_no_temp) { - auto eventProcessA = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList, - a_one, a_two, a_ld, a_offset, a_buffer, - m_ceiled, k_ceiled, m_ceiled, 0, a_temp, - ConstantOne(), program, - true, a_do_transpose, a_conjugate); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventProcessA); - } - - // As above, but now for matrix B - if (!b_no_temp) { - auto eventProcessB = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB.pointer(), emptyEventList, - b_one, b_two, b_ld, b_offset, b_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, b_temp, - ConstantOne(), program, - true, b_do_transpose, b_conjugate); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventProcessB); - } - - // As above, but now for matrix C. This is only necessary if C is used both as input and output. - if (!c_no_temp && beta != static_cast(0)) { - auto eventProcessC = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, - c_one, c_two, c_ld, c_offset, c_buffer, - m_ceiled, n_ceiled, m_ceiled, 0, c_temp, - ConstantOne(), program, - true, c_do_transpose, false); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventProcessC); - } - - // Retrieves the Xgemm kernel from the compiled binary - try { - auto kernel = Kernel(program, "Xgemm"); - - // Sets the kernel arguments - kernel.SetArgument(0, static_cast(m_ceiled)); - kernel.SetArgument(1, static_cast(n_ceiled)); - kernel.SetArgument(2, static_cast(k_ceiled)); - kernel.SetArgument(3, alpha_buffer()); - kernel.SetArgument(4, beta_buffer()); - kernel.SetArgument(5, a_temp()); - kernel.SetArgument(6, b_temp()); - kernel.SetArgument(7, c_temp()); - - // Computes the global and local thread sizes - const auto global = std::vector{ - (m_ceiled * db_["MDIMC"]) / db_["MWG"], - (n_ceiled * db_["NDIMC"]) / db_["NWG"] - }; - const auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; - - // Launches the kernel - auto eventKernel = Event(); - auto eventPointer = (!c_no_temp) ? eventKernel.pointer() : event_; - status = RunKernel(kernel, queue_, device_, global, local, eventPointer, eventWaitList); - if (ErrorIn(status)) { return status; } - - // Runs the post-processing kernel if needed - if (!c_no_temp) { - eventWaitList.push_back(eventKernel); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, - m_ceiled, n_ceiled, m_ceiled, 0, c_temp, - c_one, c_two, c_ld, c_offset, c_buffer, - ConstantOne(), program, - false, c_do_transpose, false); - if (ErrorIn(status)) { return status; } - } - - // Successfully finished the computation - return StatusCode::kSuccess; - } catch (...) { return StatusCode::kInvalidKernel; } - } catch (...) { return StatusCode::kTempBufferAllocFailure; } -} - -// ================================================================================================= - -// Compiles the templated class -template class Xgemm; -template class Xgemm; -template class Xgemm; -template class Xgemm; -template class Xgemm; - -// ================================================================================================= -} // namespace clblast diff --git a/src/routines/level3/xgemm.cpp b/src/routines/level3/xgemm.cpp new file mode 100644 index 00000000..9ea5559c --- /dev/null +++ b/src/routines/level3/xgemm.cpp @@ -0,0 +1,223 @@ + +// ================================================================================================= +// 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 Xgemm class (see the header for information about the class). +// +// ================================================================================================= + +#include "routines/level3/xgemm.hpp" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xgemm::Xgemm(Queue &queue, EventPointer event, const std::string &name): + Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue()) { + source_string_ = + #include "../../kernels/level3/level3.opencl" + #include "../../kernels/level3/copy_fast.opencl" + #include "../../kernels/level3/copy_pad.opencl" + #include "../../kernels/level3/transpose_fast.opencl" + #include "../../kernels/level3/transpose_pad.opencl" + #include "../../kernels/level3/convert_symmetric.opencl" + #include "../../kernels/level3/convert_triangular.opencl" + #include "../../kernels/level3/convert_hermitian.opencl" + #include "../../kernels/level3/xgemm_part1.opencl" + #include "../../kernels/level3/xgemm_part2.opencl" + ; +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xgemm::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 T beta, + const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { + + // Makes sure all dimensions are larger than zero + if ((m == 0) || (n == 0) || (k == 0)) { return StatusCode::kInvalidDimension; } + + // Computes whether or not the matrices are transposed in memory. This is based on their layout + // (row or column-major) and whether or not they are requested to be pre-transposed. Note + // that the Xgemm kernel expects either matrices A and C (in case of row-major) or B (in case of + // col-major) to be transformed, so transposing requirements are not the same as whether or not + // the matrix is actually transposed in memory. + const auto a_rotated = (layout == Layout::kColMajor && a_transpose != Transpose::kNo) || + (layout == Layout::kRowMajor && a_transpose == Transpose::kNo); + const auto b_rotated = (layout == Layout::kColMajor && b_transpose != Transpose::kNo) || + (layout == Layout::kRowMajor && b_transpose == Transpose::kNo); + const auto c_rotated = (layout == Layout::kRowMajor); + const auto a_do_transpose = a_rotated; + const auto b_do_transpose = !b_rotated; + const auto c_do_transpose = c_rotated; + + // In case of complex data-types, the transpose can also become a conjugate transpose + const auto a_conjugate = (a_transpose == Transpose::kConjugate); + const auto b_conjugate = (b_transpose == Transpose::kConjugate); + + // Computes the first and second dimensions of the 3 matrices taking into account whether the + // matrices are rotated or not + const auto a_one = (a_rotated) ? k : m; + const auto a_two = (a_rotated) ? m : k; + const auto b_one = (b_rotated) ? n : k; + const auto b_two = (b_rotated) ? k : n; + const auto c_one = (c_rotated) ? n : m; + const auto c_two = (c_rotated) ? m : n; + + // Tests three matrices (A, B, C) for validity, first from a perspective of the OpenCL buffers and + // their sizes, and then from a perspective of parameter values (e.g. m, n, k). Tests whether the + // OpenCL buffers are valid and non-zero and whether the OpenCL buffers have sufficient storage + // space. Also tests that the leading dimensions of: + // matrix A cannot be less than K when rotated, or less than M when not-rotated + // matrix B cannot be less than N when rotated, or less than K when not-rotated + // matrix C cannot be less than N when rotated, or less than M when not-rotated + auto status = TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld); + if (ErrorIn(status)) { return status; } + status = TestMatrixB(b_one, b_two, b_buffer, b_offset, b_ld); + if (ErrorIn(status)) { return status; } + status = TestMatrixC(c_one, c_two, c_buffer, c_offset, c_ld); + if (ErrorIn(status)) { return status; } + + // Calculates the ceiled versions of m, n, and k + const auto m_ceiled = Ceil(m, db_["MWG"]); + const auto n_ceiled = Ceil(n, db_["NWG"]); + const auto k_ceiled = Ceil(k, db_["KWG"]); + + // The padded/transposed input/output matrices: if memory allocation fails, throw an exception + try { + + // Loads the program from the database + const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); + + // Determines whether or not temporary matrices are needed + auto a_no_temp = a_one == m_ceiled && a_two == k_ceiled && a_ld == m_ceiled && a_offset == 0 && + a_do_transpose == false && a_conjugate == false; + auto b_no_temp = b_one == n_ceiled && b_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && + b_do_transpose == false && b_conjugate == false; + auto c_no_temp = c_one == m_ceiled && c_two == n_ceiled && c_ld == m_ceiled && c_offset == 0 && + c_do_transpose == false; + + // Creates the temporary matrices + const auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, k_ceiled*m_ceiled); + const auto b_temp = (b_no_temp) ? b_buffer : Buffer(context_, k_ceiled*n_ceiled); + const auto c_temp = (c_no_temp) ? c_buffer : Buffer(context_, m_ceiled*n_ceiled); + + // Upload the scalar arguments as constant buffers to the device (needed for half-precision) + auto alpha_buffer = Buffer(context_, 1); + auto beta_buffer = Buffer(context_, 1); + alpha_buffer.Write(queue_, 1, &alpha); + beta_buffer.Write(queue_, 1, &beta); + + // Events of all kernels (including pre/post processing kernels) + auto eventWaitList = std::vector(); + auto emptyEventList = std::vector(); + + // Runs the pre-processing kernel for matrix A. This transposes the matrix, but also pads zeros + // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In + // case nothing has to be done, these kernels can be skipped. + if (!a_no_temp) { + auto eventProcessA = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList, + a_one, a_two, a_ld, a_offset, a_buffer, + m_ceiled, k_ceiled, m_ceiled, 0, a_temp, + ConstantOne(), program, + true, a_do_transpose, a_conjugate); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessA); + } + + // As above, but now for matrix B + if (!b_no_temp) { + auto eventProcessB = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB.pointer(), emptyEventList, + b_one, b_two, b_ld, b_offset, b_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, b_temp, + ConstantOne(), program, + true, b_do_transpose, b_conjugate); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessB); + } + + // As above, but now for matrix C. This is only necessary if C is used both as input and output. + if (!c_no_temp && beta != static_cast(0)) { + auto eventProcessC = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, + c_one, c_two, c_ld, c_offset, c_buffer, + m_ceiled, n_ceiled, m_ceiled, 0, c_temp, + ConstantOne(), program, + true, c_do_transpose, false); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessC); + } + + // Retrieves the Xgemm kernel from the compiled binary + try { + auto kernel = Kernel(program, "Xgemm"); + + // Sets the kernel arguments + kernel.SetArgument(0, static_cast(m_ceiled)); + kernel.SetArgument(1, static_cast(n_ceiled)); + kernel.SetArgument(2, static_cast(k_ceiled)); + kernel.SetArgument(3, alpha_buffer()); + kernel.SetArgument(4, beta_buffer()); + kernel.SetArgument(5, a_temp()); + kernel.SetArgument(6, b_temp()); + kernel.SetArgument(7, c_temp()); + + // Computes the global and local thread sizes + const auto global = std::vector{ + (m_ceiled * db_["MDIMC"]) / db_["MWG"], + (n_ceiled * db_["NDIMC"]) / db_["NWG"] + }; + const auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; + + // Launches the kernel + auto eventKernel = Event(); + auto eventPointer = (!c_no_temp) ? eventKernel.pointer() : event_; + status = RunKernel(kernel, queue_, device_, global, local, eventPointer, eventWaitList); + if (ErrorIn(status)) { return status; } + + // Runs the post-processing kernel if needed + if (!c_no_temp) { + eventWaitList.push_back(eventKernel); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, + m_ceiled, n_ceiled, m_ceiled, 0, c_temp, + c_one, c_two, c_ld, c_offset, c_buffer, + ConstantOne(), program, + false, c_do_transpose, false); + if (ErrorIn(status)) { return status; } + } + + // Successfully finished the computation + return StatusCode::kSuccess; + } catch (...) { return StatusCode::kInvalidKernel; } + } catch (...) { return StatusCode::kTempBufferAllocFailure; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xgemm; +template class Xgemm; +template class Xgemm; +template class Xgemm; +template class Xgemm; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level3/xhemm.cc b/src/routines/level3/xhemm.cc deleted file mode 100644 index 9813503e..00000000 --- a/src/routines/level3/xhemm.cc +++ /dev/null @@ -1,134 +0,0 @@ - -// ================================================================================================= -// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This -// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max- -// width of 100 characters per line. -// -// Author(s): -// Cedric Nugteren -// -// This file implements the Xhemm class (see the header for information about the class). -// -// ================================================================================================= - -#include "routines/level3/xhemm.hpp" - -#include -#include - -namespace clblast { -// ================================================================================================= - -// Constructor: forwards to base class constructor -template -Xhemm::Xhemm(Queue &queue, EventPointer event, const std::string &name): - Xgemm(queue, event, name) { -} - -// ================================================================================================= - -// The main routine -template -StatusCode Xhemm::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 T beta, - const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { - - // Makes sure all dimensions are larger than zero - if ((m == 0) || (n == 0) ) { return StatusCode::kInvalidDimension; } - - // Computes the k dimension. This is based on whether or not the hermitian matrix is A (on the - // left) or B (on the right) in the Xgemm routine. - auto k = (side == Side::kLeft) ? m : n; - - // Checks for validity of the squared A matrix - auto status = TestMatrixA(k, k, a_buffer, a_offset, a_ld); - if (ErrorIn(status)) { return status; } - - // Determines which kernel to run based on the layout (the Xgemm kernel assumes column-major as - // default) and on whether we are dealing with an upper or lower triangle of the hermitian matrix - bool is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) || - (triangle == Triangle::kLower && layout == Layout::kRowMajor)); - auto kernel_name = (is_upper) ? "HermUpperToSquared" : "HermLowerToSquared"; - - // Temporary buffer for a copy of the hermitian matrix - try { - auto temp_herm = Buffer(context_, k*k); - - // Creates a general matrix from the hermitian matrix to be able to run the regular Xgemm - // routine afterwards - try { - const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); - auto kernel = Kernel(program, kernel_name); - - // Sets the arguments for the hermitian-to-squared kernel - kernel.SetArgument(0, static_cast(k)); - kernel.SetArgument(1, static_cast(a_ld)); - kernel.SetArgument(2, static_cast(a_offset)); - kernel.SetArgument(3, a_buffer()); - kernel.SetArgument(4, static_cast(k)); - kernel.SetArgument(5, static_cast(k)); - kernel.SetArgument(6, static_cast(0)); - kernel.SetArgument(7, temp_herm()); - - // Uses the common padding kernel's thread configuration. This is allowed, since the - // hermitian-to-squared kernel uses the same parameters. - auto global = std::vector{Ceil(CeilDiv(k, db_["PAD_WPTX"]), db_["PAD_DIMX"]), - Ceil(CeilDiv(k, db_["PAD_WPTY"]), db_["PAD_DIMY"])}; - auto local = std::vector{db_["PAD_DIMX"], db_["PAD_DIMY"]}; - auto kernelEvent = Event(); - status = RunKernel(kernel, queue_, device_, global, local, kernelEvent.pointer()); - if (ErrorIn(status)) { return status; } - - // Synchronize now: 'DoGemm' does not accept a list of events to wait for - kernelEvent.WaitForCompletion(); - - // Runs the regular Xgemm code with either "C := AB+C" or ... - if (side == Side::kLeft) { - status = DoGemm(layout, Transpose::kNo, Transpose::kNo, - m, n, k, - alpha, - temp_herm, 0, k, - b_buffer, b_offset, b_ld, - beta, - c_buffer, c_offset, c_ld); - } - - // ... with "C := BA+C". Note that A and B are now reversed. - else { - status = DoGemm(layout, Transpose::kNo, Transpose::kNo, - m, n, k, - alpha, - b_buffer, b_offset, b_ld, - temp_herm, 0, k, - beta, - c_buffer, c_offset, c_ld); - - // A and B are now reversed, so also reverse the error codes returned from the Xgemm routine - switch(status) { - case StatusCode::kInvalidMatrixA: status = StatusCode::kInvalidMatrixB; break; - case StatusCode::kInvalidMatrixB: status = StatusCode::kInvalidMatrixA; break; - case StatusCode::kInvalidLeadDimA: status = StatusCode::kInvalidLeadDimB; break; - case StatusCode::kInvalidLeadDimB: status = StatusCode::kInvalidLeadDimA; break; - case StatusCode::kInsufficientMemoryA: status = StatusCode::kInsufficientMemoryB; break; - case StatusCode::kInsufficientMemoryB: status = StatusCode::kInsufficientMemoryA; break; - } - } - - // Return the status of the Xgemm routine - return status; - } catch (...) { return StatusCode::kInvalidKernel; } - } catch (...) { return StatusCode::kTempBufferAllocFailure; } -} - -// ================================================================================================= - -// Compiles the templated class -template class Xhemm; -template class Xhemm; - -// ================================================================================================= -} // namespace clblast diff --git a/src/routines/level3/xhemm.cpp b/src/routines/level3/xhemm.cpp new file mode 100644 index 00000000..9813503e --- /dev/null +++ b/src/routines/level3/xhemm.cpp @@ -0,0 +1,134 @@ + +// ================================================================================================= +// 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 Xhemm class (see the header for information about the class). +// +// ================================================================================================= + +#include "routines/level3/xhemm.hpp" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xhemm::Xhemm(Queue &queue, EventPointer event, const std::string &name): + Xgemm(queue, event, name) { +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xhemm::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 T beta, + const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { + + // Makes sure all dimensions are larger than zero + if ((m == 0) || (n == 0) ) { return StatusCode::kInvalidDimension; } + + // Computes the k dimension. This is based on whether or not the hermitian matrix is A (on the + // left) or B (on the right) in the Xgemm routine. + auto k = (side == Side::kLeft) ? m : n; + + // Checks for validity of the squared A matrix + auto status = TestMatrixA(k, k, a_buffer, a_offset, a_ld); + if (ErrorIn(status)) { return status; } + + // Determines which kernel to run based on the layout (the Xgemm kernel assumes column-major as + // default) and on whether we are dealing with an upper or lower triangle of the hermitian matrix + bool is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) || + (triangle == Triangle::kLower && layout == Layout::kRowMajor)); + auto kernel_name = (is_upper) ? "HermUpperToSquared" : "HermLowerToSquared"; + + // Temporary buffer for a copy of the hermitian matrix + try { + auto temp_herm = Buffer(context_, k*k); + + // Creates a general matrix from the hermitian matrix to be able to run the regular Xgemm + // routine afterwards + try { + const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); + auto kernel = Kernel(program, kernel_name); + + // Sets the arguments for the hermitian-to-squared kernel + kernel.SetArgument(0, static_cast(k)); + kernel.SetArgument(1, static_cast(a_ld)); + kernel.SetArgument(2, static_cast(a_offset)); + kernel.SetArgument(3, a_buffer()); + kernel.SetArgument(4, static_cast(k)); + kernel.SetArgument(5, static_cast(k)); + kernel.SetArgument(6, static_cast(0)); + kernel.SetArgument(7, temp_herm()); + + // Uses the common padding kernel's thread configuration. This is allowed, since the + // hermitian-to-squared kernel uses the same parameters. + auto global = std::vector{Ceil(CeilDiv(k, db_["PAD_WPTX"]), db_["PAD_DIMX"]), + Ceil(CeilDiv(k, db_["PAD_WPTY"]), db_["PAD_DIMY"])}; + auto local = std::vector{db_["PAD_DIMX"], db_["PAD_DIMY"]}; + auto kernelEvent = Event(); + status = RunKernel(kernel, queue_, device_, global, local, kernelEvent.pointer()); + if (ErrorIn(status)) { return status; } + + // Synchronize now: 'DoGemm' does not accept a list of events to wait for + kernelEvent.WaitForCompletion(); + + // Runs the regular Xgemm code with either "C := AB+C" or ... + if (side == Side::kLeft) { + status = DoGemm(layout, Transpose::kNo, Transpose::kNo, + m, n, k, + alpha, + temp_herm, 0, k, + b_buffer, b_offset, b_ld, + beta, + c_buffer, c_offset, c_ld); + } + + // ... with "C := BA+C". Note that A and B are now reversed. + else { + status = DoGemm(layout, Transpose::kNo, Transpose::kNo, + m, n, k, + alpha, + b_buffer, b_offset, b_ld, + temp_herm, 0, k, + beta, + c_buffer, c_offset, c_ld); + + // A and B are now reversed, so also reverse the error codes returned from the Xgemm routine + switch(status) { + case StatusCode::kInvalidMatrixA: status = StatusCode::kInvalidMatrixB; break; + case StatusCode::kInvalidMatrixB: status = StatusCode::kInvalidMatrixA; break; + case StatusCode::kInvalidLeadDimA: status = StatusCode::kInvalidLeadDimB; break; + case StatusCode::kInvalidLeadDimB: status = StatusCode::kInvalidLeadDimA; break; + case StatusCode::kInsufficientMemoryA: status = StatusCode::kInsufficientMemoryB; break; + case StatusCode::kInsufficientMemoryB: status = StatusCode::kInsufficientMemoryA; break; + } + } + + // Return the status of the Xgemm routine + return status; + } catch (...) { return StatusCode::kInvalidKernel; } + } catch (...) { return StatusCode::kTempBufferAllocFailure; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xhemm; +template class Xhemm; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level3/xher2k.cc b/src/routines/level3/xher2k.cc deleted file mode 100644 index bd7a053e..00000000 --- a/src/routines/level3/xher2k.cc +++ /dev/null @@ -1,241 +0,0 @@ - -// ================================================================================================= -// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This -// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max- -// width of 100 characters per line. -// -// Author(s): -// Cedric Nugteren -// -// This file implements the Xher2k class (see the header for information about the class). -// -// ================================================================================================= - -#include "routines/level3/xher2k.hpp" - -#include -#include - -namespace clblast { -// ================================================================================================= - -// Constructor: forwards to base class constructor -template -Xher2k::Xher2k(Queue &queue, EventPointer event, const std::string &name): - Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue()) { - source_string_ = - #include "../../kernels/level3/level3.opencl" - #include "../../kernels/level3/copy_fast.opencl" - #include "../../kernels/level3/copy_pad.opencl" - #include "../../kernels/level3/transpose_fast.opencl" - #include "../../kernels/level3/transpose_pad.opencl" - #include "../../kernels/level3/xgemm_part1.opencl" - #include "../../kernels/level3/xgemm_part2.opencl" - ; -} - -// ================================================================================================= - -// The main routine -template -StatusCode Xher2k::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 U beta, - const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { - - // Makes sure all dimensions are larger than zero - if ((n == 0) || (k == 0) ) { return StatusCode::kInvalidDimension; } - - // Determines whether to apply the conjugate transpose to matrix B (argument: no transpose) or - // to matrix A (argument: conjugate transpose) - auto ab_conjugate = (ab_transpose != Transpose::kNo); - - // Computes whether or not the matrices are transposed in memory. This is based on their layout - // (row or column-major) and whether or not they are requested to be pre-transposed. - auto ab_rotated = (layout == Layout::kColMajor && ab_conjugate) || - (layout == Layout::kRowMajor && !ab_conjugate); - auto c_rotated = (layout == Layout::kRowMajor); - - // Computes the first and second dimensions of the A and B matrices taking the layout into account - auto ab_one = (ab_rotated) ? k : n; - auto ab_two = (ab_rotated) ? n : k; - - // Tests the matrices (A, B, C) for validity, first from a perspective of the OpenCL buffers and - // their sizes, and then from a perspective of parameter values (e.g. n, k). Tests whether the - // OpenCL buffers are valid and non-zero and whether the OpenCL buffers have sufficient storage - // space. Also tests that the leading dimensions of: - // matrix A cannot be less than N when rotated, or less than K when not-rotated - // matrix B cannot be less than N when rotated, or less than K when not-rotated - // matrix C cannot be less than N - auto status = TestMatrixA(ab_one, ab_two, a_buffer, a_offset, a_ld); - if (ErrorIn(status)) { return status; } - status = TestMatrixB(ab_one, ab_two, b_buffer, b_offset, b_ld); - if (ErrorIn(status)) { return status; } - status = TestMatrixC(n, n, c_buffer, c_offset, c_ld); - if (ErrorIn(status)) { return status; } - - // Calculates the ceiled versions of n and k - auto n_ceiled = Ceil(n, db_["NWG"]); - auto k_ceiled = Ceil(k, db_["KWG"]); - - // Decides which kernel to run: the upper-triangular or lower-triangular version - auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower"; - - // The padded/transposed input/output matrices: if memory allocation fails, throw an exception - try { - - // Loads the program from the database - const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); - - // Determines whether or not temporary matrices are needed - auto a1_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && - ab_rotated == false && ab_conjugate == false; - auto a2_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && - ab_rotated == false && ab_conjugate == true; - auto b1_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && - ab_rotated == false && ab_conjugate == false; - auto b2_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && - ab_rotated == false && ab_conjugate == true; - - // Creates the temporary matrices - auto a1_temp = (a1_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto a2_temp = (a2_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto b1_temp = (b1_no_temp) ? b_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto b2_temp = (b2_no_temp) ? b_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto c_temp = Buffer(context_, n_ceiled*n_ceiled); - - // Upload the scalar arguments as constant buffers to the device (needed for half-precision) - auto complex_beta = T{beta, static_cast(0.0)}; - auto alpha_buffer = Buffer(context_, 1); - auto beta_buffer = Buffer(context_, 1); - alpha_buffer.Write(queue_, 1, &alpha); - beta_buffer.Write(queue_, 1, &complex_beta); - - // Events of all kernels (including pre/post processing kernels) - auto eventWaitList = std::vector(); - auto emptyEventList = std::vector(); - - // Runs the pre-processing kernels. This transposes the matrices A and B, but also pads zeros to - // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In - // case nothing has to be done, these kernels can be skipped. - if (!a1_no_temp) { - auto eventProcessA1 = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA1.pointer(), emptyEventList, - ab_one, ab_two, a_ld, a_offset, a_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, a1_temp, - ConstantOne(), program, - true, ab_rotated, ab_conjugate); - eventWaitList.push_back(eventProcessA1); - if (ErrorIn(status)) { return status; } - } - if (!a2_no_temp) { - auto eventProcessA2 = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA2.pointer(), emptyEventList, - ab_one, ab_two, a_ld, a_offset, a_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, a2_temp, - ConstantOne(), program, - true, ab_rotated, !ab_conjugate); - eventWaitList.push_back(eventProcessA2); - if (ErrorIn(status)) { return status; } - } - if (!b1_no_temp) { - auto eventProcessB1 = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB1.pointer(), emptyEventList, - ab_one, ab_two, b_ld, b_offset, b_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, b1_temp, - ConstantOne(), program, - true, ab_rotated, ab_conjugate); - eventWaitList.push_back(eventProcessB1); - if (ErrorIn(status)) { return status; } - } - if (!b2_no_temp) { - auto eventProcessB2 = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB2.pointer(), emptyEventList, - ab_one, ab_two, b_ld, b_offset, b_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, b2_temp, - ConstantOne(), program, - true, ab_rotated, !ab_conjugate); - eventWaitList.push_back(eventProcessB2); - if (ErrorIn(status)) { return status; } - } - - // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to - // modify the other triangle. - auto eventProcessC = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, - n, n, c_ld, c_offset, c_buffer, - n_ceiled, n_ceiled, n_ceiled, 0, c_temp, - ConstantOne(), program, - true, c_rotated, false); - eventWaitList.push_back(eventProcessC); - if (ErrorIn(status)) { return status; } - - // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary - try { - auto kernel = Kernel(program, kernel_name); - - // Sets the kernel arguments - kernel.SetArgument(0, static_cast(n_ceiled)); - kernel.SetArgument(1, static_cast(k_ceiled)); - kernel.SetArgument(2, alpha_buffer()); - kernel.SetArgument(3, beta_buffer()); - kernel.SetArgument(4, a1_temp()); - kernel.SetArgument(5, b2_temp()); - kernel.SetArgument(6, c_temp()); - - // Computes the global and local thread sizes - auto global = std::vector{ - (n_ceiled * db_["MDIMC"]) / db_["MWG"], - (n_ceiled * db_["NDIMC"]) / db_["NWG"] - }; - auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; - - // Launches the kernel - auto eventKernel1 = Event(); - status = RunKernel(kernel, queue_, device_, global, local, eventKernel1.pointer(), eventWaitList); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventKernel1); - - // Swaps the arguments for matrices A and B, sets 'beta' to 1, and conjugate alpha - auto conjugate_alpha = T{alpha.real(), -alpha.imag()}; - auto complex_one = T{static_cast(1.0), static_cast(0.0)}; - alpha_buffer.Write(queue_, 1, &conjugate_alpha); - beta_buffer.Write(queue_, 1, &complex_one); - kernel.SetArgument(2, alpha_buffer()); - kernel.SetArgument(3, beta_buffer()); - kernel.SetArgument(4, b1_temp()); - kernel.SetArgument(5, a2_temp()); - - // Runs the kernel again - auto eventKernel2 = Event(); - status = RunKernel(kernel, queue_, device_, global, local, eventKernel2.pointer(), eventWaitList); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventKernel2); - - // Runs the post-processing kernel - auto upper = (triangle == Triangle::kUpper); - auto lower = (triangle == Triangle::kLower); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, - n_ceiled, n_ceiled, n_ceiled, 0, c_temp, - n, n, c_ld, c_offset, c_buffer, - ConstantOne(), program, - false, c_rotated, false, upper, lower, true); - if (ErrorIn(status)) { return status; } - - // Successfully finished the computation - return StatusCode::kSuccess; - } catch (...) { return StatusCode::kInvalidKernel; } - } catch (...) { return StatusCode::kTempBufferAllocFailure; } -} - -// ================================================================================================= - -// Compiles the templated class -template class Xher2k; -template class Xher2k; - -// ================================================================================================= -} // namespace clblast diff --git a/src/routines/level3/xher2k.cpp b/src/routines/level3/xher2k.cpp new file mode 100644 index 00000000..bd7a053e --- /dev/null +++ b/src/routines/level3/xher2k.cpp @@ -0,0 +1,241 @@ + +// ================================================================================================= +// 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 Xher2k class (see the header for information about the class). +// +// ================================================================================================= + +#include "routines/level3/xher2k.hpp" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xher2k::Xher2k(Queue &queue, EventPointer event, const std::string &name): + Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue()) { + source_string_ = + #include "../../kernels/level3/level3.opencl" + #include "../../kernels/level3/copy_fast.opencl" + #include "../../kernels/level3/copy_pad.opencl" + #include "../../kernels/level3/transpose_fast.opencl" + #include "../../kernels/level3/transpose_pad.opencl" + #include "../../kernels/level3/xgemm_part1.opencl" + #include "../../kernels/level3/xgemm_part2.opencl" + ; +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xher2k::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 U beta, + const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { + + // Makes sure all dimensions are larger than zero + if ((n == 0) || (k == 0) ) { return StatusCode::kInvalidDimension; } + + // Determines whether to apply the conjugate transpose to matrix B (argument: no transpose) or + // to matrix A (argument: conjugate transpose) + auto ab_conjugate = (ab_transpose != Transpose::kNo); + + // Computes whether or not the matrices are transposed in memory. This is based on their layout + // (row or column-major) and whether or not they are requested to be pre-transposed. + auto ab_rotated = (layout == Layout::kColMajor && ab_conjugate) || + (layout == Layout::kRowMajor && !ab_conjugate); + auto c_rotated = (layout == Layout::kRowMajor); + + // Computes the first and second dimensions of the A and B matrices taking the layout into account + auto ab_one = (ab_rotated) ? k : n; + auto ab_two = (ab_rotated) ? n : k; + + // Tests the matrices (A, B, C) for validity, first from a perspective of the OpenCL buffers and + // their sizes, and then from a perspective of parameter values (e.g. n, k). Tests whether the + // OpenCL buffers are valid and non-zero and whether the OpenCL buffers have sufficient storage + // space. Also tests that the leading dimensions of: + // matrix A cannot be less than N when rotated, or less than K when not-rotated + // matrix B cannot be less than N when rotated, or less than K when not-rotated + // matrix C cannot be less than N + auto status = TestMatrixA(ab_one, ab_two, a_buffer, a_offset, a_ld); + if (ErrorIn(status)) { return status; } + status = TestMatrixB(ab_one, ab_two, b_buffer, b_offset, b_ld); + if (ErrorIn(status)) { return status; } + status = TestMatrixC(n, n, c_buffer, c_offset, c_ld); + if (ErrorIn(status)) { return status; } + + // Calculates the ceiled versions of n and k + auto n_ceiled = Ceil(n, db_["NWG"]); + auto k_ceiled = Ceil(k, db_["KWG"]); + + // Decides which kernel to run: the upper-triangular or lower-triangular version + auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower"; + + // The padded/transposed input/output matrices: if memory allocation fails, throw an exception + try { + + // Loads the program from the database + const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); + + // Determines whether or not temporary matrices are needed + auto a1_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && + ab_rotated == false && ab_conjugate == false; + auto a2_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && + ab_rotated == false && ab_conjugate == true; + auto b1_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && + ab_rotated == false && ab_conjugate == false; + auto b2_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && + ab_rotated == false && ab_conjugate == true; + + // Creates the temporary matrices + auto a1_temp = (a1_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); + auto a2_temp = (a2_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); + auto b1_temp = (b1_no_temp) ? b_buffer : Buffer(context_, k_ceiled*n_ceiled); + auto b2_temp = (b2_no_temp) ? b_buffer : Buffer(context_, k_ceiled*n_ceiled); + auto c_temp = Buffer(context_, n_ceiled*n_ceiled); + + // Upload the scalar arguments as constant buffers to the device (needed for half-precision) + auto complex_beta = T{beta, static_cast(0.0)}; + auto alpha_buffer = Buffer(context_, 1); + auto beta_buffer = Buffer(context_, 1); + alpha_buffer.Write(queue_, 1, &alpha); + beta_buffer.Write(queue_, 1, &complex_beta); + + // Events of all kernels (including pre/post processing kernels) + auto eventWaitList = std::vector(); + auto emptyEventList = std::vector(); + + // Runs the pre-processing kernels. This transposes the matrices A and B, but also pads zeros to + // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In + // case nothing has to be done, these kernels can be skipped. + if (!a1_no_temp) { + auto eventProcessA1 = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA1.pointer(), emptyEventList, + ab_one, ab_two, a_ld, a_offset, a_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, a1_temp, + ConstantOne(), program, + true, ab_rotated, ab_conjugate); + eventWaitList.push_back(eventProcessA1); + if (ErrorIn(status)) { return status; } + } + if (!a2_no_temp) { + auto eventProcessA2 = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA2.pointer(), emptyEventList, + ab_one, ab_two, a_ld, a_offset, a_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, a2_temp, + ConstantOne(), program, + true, ab_rotated, !ab_conjugate); + eventWaitList.push_back(eventProcessA2); + if (ErrorIn(status)) { return status; } + } + if (!b1_no_temp) { + auto eventProcessB1 = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB1.pointer(), emptyEventList, + ab_one, ab_two, b_ld, b_offset, b_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, b1_temp, + ConstantOne(), program, + true, ab_rotated, ab_conjugate); + eventWaitList.push_back(eventProcessB1); + if (ErrorIn(status)) { return status; } + } + if (!b2_no_temp) { + auto eventProcessB2 = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB2.pointer(), emptyEventList, + ab_one, ab_two, b_ld, b_offset, b_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, b2_temp, + ConstantOne(), program, + true, ab_rotated, !ab_conjugate); + eventWaitList.push_back(eventProcessB2); + if (ErrorIn(status)) { return status; } + } + + // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to + // modify the other triangle. + auto eventProcessC = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, + n, n, c_ld, c_offset, c_buffer, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + ConstantOne(), program, + true, c_rotated, false); + eventWaitList.push_back(eventProcessC); + if (ErrorIn(status)) { return status; } + + // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary + try { + auto kernel = Kernel(program, kernel_name); + + // Sets the kernel arguments + kernel.SetArgument(0, static_cast(n_ceiled)); + kernel.SetArgument(1, static_cast(k_ceiled)); + kernel.SetArgument(2, alpha_buffer()); + kernel.SetArgument(3, beta_buffer()); + kernel.SetArgument(4, a1_temp()); + kernel.SetArgument(5, b2_temp()); + kernel.SetArgument(6, c_temp()); + + // Computes the global and local thread sizes + auto global = std::vector{ + (n_ceiled * db_["MDIMC"]) / db_["MWG"], + (n_ceiled * db_["NDIMC"]) / db_["NWG"] + }; + auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; + + // Launches the kernel + auto eventKernel1 = Event(); + status = RunKernel(kernel, queue_, device_, global, local, eventKernel1.pointer(), eventWaitList); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventKernel1); + + // Swaps the arguments for matrices A and B, sets 'beta' to 1, and conjugate alpha + auto conjugate_alpha = T{alpha.real(), -alpha.imag()}; + auto complex_one = T{static_cast(1.0), static_cast(0.0)}; + alpha_buffer.Write(queue_, 1, &conjugate_alpha); + beta_buffer.Write(queue_, 1, &complex_one); + kernel.SetArgument(2, alpha_buffer()); + kernel.SetArgument(3, beta_buffer()); + kernel.SetArgument(4, b1_temp()); + kernel.SetArgument(5, a2_temp()); + + // Runs the kernel again + auto eventKernel2 = Event(); + status = RunKernel(kernel, queue_, device_, global, local, eventKernel2.pointer(), eventWaitList); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventKernel2); + + // Runs the post-processing kernel + auto upper = (triangle == Triangle::kUpper); + auto lower = (triangle == Triangle::kLower); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + n, n, c_ld, c_offset, c_buffer, + ConstantOne(), program, + false, c_rotated, false, upper, lower, true); + if (ErrorIn(status)) { return status; } + + // Successfully finished the computation + return StatusCode::kSuccess; + } catch (...) { return StatusCode::kInvalidKernel; } + } catch (...) { return StatusCode::kTempBufferAllocFailure; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xher2k; +template class Xher2k; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level3/xherk.cc b/src/routines/level3/xherk.cc deleted file mode 100644 index 6ef7f21f..00000000 --- a/src/routines/level3/xherk.cc +++ /dev/null @@ -1,197 +0,0 @@ - -// ================================================================================================= -// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This -// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max- -// width of 100 characters per line. -// -// Author(s): -// Cedric Nugteren -// -// This file implements the Xherk class (see the header for information about the class). -// -// ================================================================================================= - -#include "routines/level3/xherk.hpp" - -#include -#include - -namespace clblast { -// ================================================================================================= - -// Constructor: forwards to base class constructor -template -Xherk::Xherk(Queue &queue, EventPointer event, const std::string &name): - Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue()) { - source_string_ = - #include "../../kernels/level3/level3.opencl" - #include "../../kernels/level3/copy_fast.opencl" - #include "../../kernels/level3/copy_pad.opencl" - #include "../../kernels/level3/transpose_fast.opencl" - #include "../../kernels/level3/transpose_pad.opencl" - #include "../../kernels/level3/xgemm_part1.opencl" - #include "../../kernels/level3/xgemm_part2.opencl" - ; -} - -// ================================================================================================= - -// The main routine -template -StatusCode Xherk::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 U beta, - const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { - - // Makes sure all dimensions are larger than zero - if ((n == 0) || (k == 0) ) { return StatusCode::kInvalidDimension; } - - // Determines whether to apply the conjugate transpose to matrix B (argument: no transpose) or - // to matrix A (argument: conjugate transpose) - auto a_conjugate = (a_transpose != Transpose::kNo); - auto b_conjugate = (a_transpose == Transpose::kNo); - - // Computes whether or not the matrices are transposed in memory. This is based on their layout - // (row or column-major) and whether or not they are requested to be pre-transposed. - auto a_rotated = (layout == Layout::kColMajor && a_conjugate) || - (layout == Layout::kRowMajor && !a_conjugate); - auto c_rotated = (layout == Layout::kRowMajor); - - // Computes the first and second dimensions of the A matrix taking the layout into account - auto a_one = (a_rotated) ? k : n; - auto a_two = (a_rotated) ? n : k; - - // Tests the two matrices (A, C) for validity, first from a perspective of the OpenCL buffers and - // their sizes, and then from a perspective of parameter values (e.g. n, k). Tests whether the - // OpenCL buffers are valid and non-zero and whether the OpenCL buffers have sufficient storage - // space. Also tests that the leading dimensions of: - // matrix A cannot be less than N when rotated, or less than K when not-rotated - // matrix C cannot be less than N - auto status = TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld); - if (ErrorIn(status)) { return status; } - status = TestMatrixC(n, n, c_buffer, c_offset, c_ld); - if (ErrorIn(status)) { return status; } - - // Calculates the ceiled versions of n and k - auto n_ceiled = Ceil(n, db_["NWG"]); - auto k_ceiled = Ceil(k, db_["KWG"]); - - // Decides which kernel to run: the upper-triangular or lower-triangular version - auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower"; - - // The padded/transposed input/output matrices: if memory allocation fails, throw an exception - try { - - // Loads the program from the database - const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); - - // Determines whether or not temporary matrices are needed - auto a_no_temp = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && - a_rotated == false && a_conjugate == false; - auto b_no_temp = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && - a_rotated == false && b_conjugate == false; - - // Creates the temporary matrices - auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto b_temp = (b_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto c_temp = Buffer(context_, n_ceiled*n_ceiled); - - // Upload the scalar arguments as constant buffers to the device (needed for half-precision) - auto complex_alpha = T{alpha, static_cast(0.0)}; - auto complex_beta = T{beta, static_cast(0.0)}; - auto alpha_buffer = Buffer(context_, 1); - auto beta_buffer = Buffer(context_, 1); - alpha_buffer.Write(queue_, 1, &complex_alpha); - beta_buffer.Write(queue_, 1, &complex_beta); - - // Events of all kernels (including pre/post processing kernels) - auto eventWaitList = std::vector(); - auto emptyEventList = std::vector(); - - // Runs the pre-processing kernel for matrix A. This transposes the matrix, but also pads zeros - // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In - // case nothing has to be done, these kernels can be skipped. Two copies are created. - if (!a_no_temp) { - auto eventProcessA = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList, - a_one, a_two, a_ld, a_offset, a_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, a_temp, - ConstantOne(), program, - true, a_rotated, a_conjugate); - eventWaitList.push_back(eventProcessA); - if (ErrorIn(status)) { return status; } - } - if (!b_no_temp) { - auto eventProcessB = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB.pointer(), emptyEventList, - a_one, a_two, a_ld, a_offset, a_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, b_temp, - ConstantOne(), program, - true, a_rotated, b_conjugate); - eventWaitList.push_back(eventProcessB); - if (ErrorIn(status)) { return status; } - } - - // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to - // modify the other triangle. - auto eventProcessC = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, - n, n, c_ld, c_offset, c_buffer, - n_ceiled, n_ceiled, n_ceiled, 0, c_temp, - ConstantOne(), program, - true, c_rotated, false); - eventWaitList.push_back(eventProcessC); - if (ErrorIn(status)) { return status; } - - // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary - try { - auto kernel = Kernel(program, kernel_name); - - // Sets the kernel arguments - kernel.SetArgument(0, static_cast(n_ceiled)); - kernel.SetArgument(1, static_cast(k_ceiled)); - kernel.SetArgument(2, alpha_buffer()); - kernel.SetArgument(3, beta_buffer()); - kernel.SetArgument(4, a_temp()); - kernel.SetArgument(5, b_temp()); - kernel.SetArgument(6, c_temp()); - - // Computes the global and local thread sizes - auto global = std::vector{ - (n_ceiled * db_["MDIMC"]) / db_["MWG"], - (n_ceiled * db_["NDIMC"]) / db_["NWG"] - }; - auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; - - // Launches the kernel - auto eventKernel = Event(); - status = RunKernel(kernel, queue_, device_, global, local, eventKernel.pointer(), eventWaitList); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventKernel); - - // Runs the post-processing kernel - auto upper = (triangle == Triangle::kUpper); - auto lower = (triangle == Triangle::kLower); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, - n_ceiled, n_ceiled, n_ceiled, 0, c_temp, - n, n, c_ld, c_offset, c_buffer, - ConstantOne(), program, - false, c_rotated, false, upper, lower, true); - if (ErrorIn(status)) { return status; } - - // Successfully finished the computation - return StatusCode::kSuccess; - } catch (...) { return StatusCode::kInvalidKernel; } - } catch (...) { return StatusCode::kTempBufferAllocFailure; } -} - -// ================================================================================================= - -// Compiles the templated class -template class Xherk; -template class Xherk; - -// ================================================================================================= -} // namespace clblast diff --git a/src/routines/level3/xherk.cpp b/src/routines/level3/xherk.cpp new file mode 100644 index 00000000..6ef7f21f --- /dev/null +++ b/src/routines/level3/xherk.cpp @@ -0,0 +1,197 @@ + +// ================================================================================================= +// 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 Xherk class (see the header for information about the class). +// +// ================================================================================================= + +#include "routines/level3/xherk.hpp" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xherk::Xherk(Queue &queue, EventPointer event, const std::string &name): + Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue()) { + source_string_ = + #include "../../kernels/level3/level3.opencl" + #include "../../kernels/level3/copy_fast.opencl" + #include "../../kernels/level3/copy_pad.opencl" + #include "../../kernels/level3/transpose_fast.opencl" + #include "../../kernels/level3/transpose_pad.opencl" + #include "../../kernels/level3/xgemm_part1.opencl" + #include "../../kernels/level3/xgemm_part2.opencl" + ; +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xherk::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 U beta, + const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { + + // Makes sure all dimensions are larger than zero + if ((n == 0) || (k == 0) ) { return StatusCode::kInvalidDimension; } + + // Determines whether to apply the conjugate transpose to matrix B (argument: no transpose) or + // to matrix A (argument: conjugate transpose) + auto a_conjugate = (a_transpose != Transpose::kNo); + auto b_conjugate = (a_transpose == Transpose::kNo); + + // Computes whether or not the matrices are transposed in memory. This is based on their layout + // (row or column-major) and whether or not they are requested to be pre-transposed. + auto a_rotated = (layout == Layout::kColMajor && a_conjugate) || + (layout == Layout::kRowMajor && !a_conjugate); + auto c_rotated = (layout == Layout::kRowMajor); + + // Computes the first and second dimensions of the A matrix taking the layout into account + auto a_one = (a_rotated) ? k : n; + auto a_two = (a_rotated) ? n : k; + + // Tests the two matrices (A, C) for validity, first from a perspective of the OpenCL buffers and + // their sizes, and then from a perspective of parameter values (e.g. n, k). Tests whether the + // OpenCL buffers are valid and non-zero and whether the OpenCL buffers have sufficient storage + // space. Also tests that the leading dimensions of: + // matrix A cannot be less than N when rotated, or less than K when not-rotated + // matrix C cannot be less than N + auto status = TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld); + if (ErrorIn(status)) { return status; } + status = TestMatrixC(n, n, c_buffer, c_offset, c_ld); + if (ErrorIn(status)) { return status; } + + // Calculates the ceiled versions of n and k + auto n_ceiled = Ceil(n, db_["NWG"]); + auto k_ceiled = Ceil(k, db_["KWG"]); + + // Decides which kernel to run: the upper-triangular or lower-triangular version + auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower"; + + // The padded/transposed input/output matrices: if memory allocation fails, throw an exception + try { + + // Loads the program from the database + const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); + + // Determines whether or not temporary matrices are needed + auto a_no_temp = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && + a_rotated == false && a_conjugate == false; + auto b_no_temp = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && + a_rotated == false && b_conjugate == false; + + // Creates the temporary matrices + auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); + auto b_temp = (b_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); + auto c_temp = Buffer(context_, n_ceiled*n_ceiled); + + // Upload the scalar arguments as constant buffers to the device (needed for half-precision) + auto complex_alpha = T{alpha, static_cast(0.0)}; + auto complex_beta = T{beta, static_cast(0.0)}; + auto alpha_buffer = Buffer(context_, 1); + auto beta_buffer = Buffer(context_, 1); + alpha_buffer.Write(queue_, 1, &complex_alpha); + beta_buffer.Write(queue_, 1, &complex_beta); + + // Events of all kernels (including pre/post processing kernels) + auto eventWaitList = std::vector(); + auto emptyEventList = std::vector(); + + // Runs the pre-processing kernel for matrix A. This transposes the matrix, but also pads zeros + // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In + // case nothing has to be done, these kernels can be skipped. Two copies are created. + if (!a_no_temp) { + auto eventProcessA = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList, + a_one, a_two, a_ld, a_offset, a_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, a_temp, + ConstantOne(), program, + true, a_rotated, a_conjugate); + eventWaitList.push_back(eventProcessA); + if (ErrorIn(status)) { return status; } + } + if (!b_no_temp) { + auto eventProcessB = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB.pointer(), emptyEventList, + a_one, a_two, a_ld, a_offset, a_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, b_temp, + ConstantOne(), program, + true, a_rotated, b_conjugate); + eventWaitList.push_back(eventProcessB); + if (ErrorIn(status)) { return status; } + } + + // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to + // modify the other triangle. + auto eventProcessC = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, + n, n, c_ld, c_offset, c_buffer, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + ConstantOne(), program, + true, c_rotated, false); + eventWaitList.push_back(eventProcessC); + if (ErrorIn(status)) { return status; } + + // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary + try { + auto kernel = Kernel(program, kernel_name); + + // Sets the kernel arguments + kernel.SetArgument(0, static_cast(n_ceiled)); + kernel.SetArgument(1, static_cast(k_ceiled)); + kernel.SetArgument(2, alpha_buffer()); + kernel.SetArgument(3, beta_buffer()); + kernel.SetArgument(4, a_temp()); + kernel.SetArgument(5, b_temp()); + kernel.SetArgument(6, c_temp()); + + // Computes the global and local thread sizes + auto global = std::vector{ + (n_ceiled * db_["MDIMC"]) / db_["MWG"], + (n_ceiled * db_["NDIMC"]) / db_["NWG"] + }; + auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; + + // Launches the kernel + auto eventKernel = Event(); + status = RunKernel(kernel, queue_, device_, global, local, eventKernel.pointer(), eventWaitList); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventKernel); + + // Runs the post-processing kernel + auto upper = (triangle == Triangle::kUpper); + auto lower = (triangle == Triangle::kLower); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + n, n, c_ld, c_offset, c_buffer, + ConstantOne(), program, + false, c_rotated, false, upper, lower, true); + if (ErrorIn(status)) { return status; } + + // Successfully finished the computation + return StatusCode::kSuccess; + } catch (...) { return StatusCode::kInvalidKernel; } + } catch (...) { return StatusCode::kTempBufferAllocFailure; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xherk; +template class Xherk; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level3/xsymm.cc b/src/routines/level3/xsymm.cc deleted file mode 100644 index 04e4b718..00000000 --- a/src/routines/level3/xsymm.cc +++ /dev/null @@ -1,137 +0,0 @@ - -// ================================================================================================= -// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This -// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max- -// width of 100 characters per line. -// -// Author(s): -// Cedric Nugteren -// -// This file implements the Xsymm class (see the header for information about the class). -// -// ================================================================================================= - -#include "routines/level3/xsymm.hpp" - -#include -#include - -namespace clblast { -// ================================================================================================= - -// Constructor: forwards to base class constructor -template -Xsymm::Xsymm(Queue &queue, EventPointer event, const std::string &name): - Xgemm(queue, event, name) { -} - -// ================================================================================================= - -// The main routine -template -StatusCode Xsymm::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 T beta, - const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { - - // Makes sure all dimensions are larger than zero - if ((m == 0) || (n == 0) ) { return StatusCode::kInvalidDimension; } - - // Computes the k dimension. This is based on whether or not the symmetric matrix is A (on the - // left) or B (on the right) in the Xgemm routine. - auto k = (side == Side::kLeft) ? m : n; - - // Checks for validity of the squared A matrix - auto status = TestMatrixA(k, k, a_buffer, a_offset, a_ld); - if (ErrorIn(status)) { return status; } - - // Determines which kernel to run based on the layout (the Xgemm kernel assumes column-major as - // default) and on whether we are dealing with an upper or lower triangle of the symmetric matrix - bool is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) || - (triangle == Triangle::kLower && layout == Layout::kRowMajor)); - auto kernel_name = (is_upper) ? "SymmUpperToSquared" : "SymmLowerToSquared"; - - // Temporary buffer for a copy of the symmetric matrix - try { - auto temp_symm = Buffer(context_, k*k); - - // Creates a general matrix from the symmetric matrix to be able to run the regular Xgemm - // routine afterwards - try { - const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); - auto kernel = Kernel(program, kernel_name); - - // Sets the arguments for the symmetric-to-squared kernel - kernel.SetArgument(0, static_cast(k)); - kernel.SetArgument(1, static_cast(a_ld)); - kernel.SetArgument(2, static_cast(a_offset)); - kernel.SetArgument(3, a_buffer()); - kernel.SetArgument(4, static_cast(k)); - kernel.SetArgument(5, static_cast(k)); - kernel.SetArgument(6, static_cast(0)); - kernel.SetArgument(7, temp_symm()); - - // Uses the common padding kernel's thread configuration. This is allowed, since the - // symmetric-to-squared kernel uses the same parameters. - auto global = std::vector{Ceil(CeilDiv(k, db_["PAD_WPTX"]), db_["PAD_DIMX"]), - Ceil(CeilDiv(k, db_["PAD_WPTY"]), db_["PAD_DIMY"])}; - auto local = std::vector{db_["PAD_DIMX"], db_["PAD_DIMY"]}; - auto kernelEvent = Event(); - status = RunKernel(kernel, queue_, device_, global, local, kernelEvent.pointer()); - if (ErrorIn(status)) { return status; } - - // Synchronize now: 'DoGemm' does not accept a list of events to wait for - kernelEvent.WaitForCompletion(); - - // Runs the regular Xgemm code with either "C := AB+C" or ... - if (side == Side::kLeft) { - status = DoGemm(layout, Transpose::kNo, Transpose::kNo, - m, n, k, - alpha, - temp_symm, 0, k, - b_buffer, b_offset, b_ld, - beta, - c_buffer, c_offset, c_ld); - } - - // ... with "C := BA+C". Note that A and B are now reversed. - else { - status = DoGemm(layout, Transpose::kNo, Transpose::kNo, - m, n, k, - alpha, - b_buffer, b_offset, b_ld, - temp_symm, 0, k, - beta, - c_buffer, c_offset, c_ld); - - // A and B are now reversed, so also reverse the error codes returned from the Xgemm routine - switch(status) { - case StatusCode::kInvalidMatrixA: status = StatusCode::kInvalidMatrixB; break; - case StatusCode::kInvalidMatrixB: status = StatusCode::kInvalidMatrixA; break; - case StatusCode::kInvalidLeadDimA: status = StatusCode::kInvalidLeadDimB; break; - case StatusCode::kInvalidLeadDimB: status = StatusCode::kInvalidLeadDimA; break; - case StatusCode::kInsufficientMemoryA: status = StatusCode::kInsufficientMemoryB; break; - case StatusCode::kInsufficientMemoryB: status = StatusCode::kInsufficientMemoryA; break; - } - } - - // Return the status of the Xgemm routine - return status; - } catch (...) { return StatusCode::kInvalidKernel; } - } catch (...) { return StatusCode::kTempBufferAllocFailure; } -} - -// ================================================================================================= - -// Compiles the templated class -template class Xsymm; -template class Xsymm; -template class Xsymm; -template class Xsymm; -template class Xsymm; - -// ================================================================================================= -} // namespace clblast diff --git a/src/routines/level3/xsymm.cpp b/src/routines/level3/xsymm.cpp new file mode 100644 index 00000000..04e4b718 --- /dev/null +++ b/src/routines/level3/xsymm.cpp @@ -0,0 +1,137 @@ + +// ================================================================================================= +// 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 Xsymm class (see the header for information about the class). +// +// ================================================================================================= + +#include "routines/level3/xsymm.hpp" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xsymm::Xsymm(Queue &queue, EventPointer event, const std::string &name): + Xgemm(queue, event, name) { +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xsymm::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 T beta, + const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { + + // Makes sure all dimensions are larger than zero + if ((m == 0) || (n == 0) ) { return StatusCode::kInvalidDimension; } + + // Computes the k dimension. This is based on whether or not the symmetric matrix is A (on the + // left) or B (on the right) in the Xgemm routine. + auto k = (side == Side::kLeft) ? m : n; + + // Checks for validity of the squared A matrix + auto status = TestMatrixA(k, k, a_buffer, a_offset, a_ld); + if (ErrorIn(status)) { return status; } + + // Determines which kernel to run based on the layout (the Xgemm kernel assumes column-major as + // default) and on whether we are dealing with an upper or lower triangle of the symmetric matrix + bool is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) || + (triangle == Triangle::kLower && layout == Layout::kRowMajor)); + auto kernel_name = (is_upper) ? "SymmUpperToSquared" : "SymmLowerToSquared"; + + // Temporary buffer for a copy of the symmetric matrix + try { + auto temp_symm = Buffer(context_, k*k); + + // Creates a general matrix from the symmetric matrix to be able to run the regular Xgemm + // routine afterwards + try { + const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); + auto kernel = Kernel(program, kernel_name); + + // Sets the arguments for the symmetric-to-squared kernel + kernel.SetArgument(0, static_cast(k)); + kernel.SetArgument(1, static_cast(a_ld)); + kernel.SetArgument(2, static_cast(a_offset)); + kernel.SetArgument(3, a_buffer()); + kernel.SetArgument(4, static_cast(k)); + kernel.SetArgument(5, static_cast(k)); + kernel.SetArgument(6, static_cast(0)); + kernel.SetArgument(7, temp_symm()); + + // Uses the common padding kernel's thread configuration. This is allowed, since the + // symmetric-to-squared kernel uses the same parameters. + auto global = std::vector{Ceil(CeilDiv(k, db_["PAD_WPTX"]), db_["PAD_DIMX"]), + Ceil(CeilDiv(k, db_["PAD_WPTY"]), db_["PAD_DIMY"])}; + auto local = std::vector{db_["PAD_DIMX"], db_["PAD_DIMY"]}; + auto kernelEvent = Event(); + status = RunKernel(kernel, queue_, device_, global, local, kernelEvent.pointer()); + if (ErrorIn(status)) { return status; } + + // Synchronize now: 'DoGemm' does not accept a list of events to wait for + kernelEvent.WaitForCompletion(); + + // Runs the regular Xgemm code with either "C := AB+C" or ... + if (side == Side::kLeft) { + status = DoGemm(layout, Transpose::kNo, Transpose::kNo, + m, n, k, + alpha, + temp_symm, 0, k, + b_buffer, b_offset, b_ld, + beta, + c_buffer, c_offset, c_ld); + } + + // ... with "C := BA+C". Note that A and B are now reversed. + else { + status = DoGemm(layout, Transpose::kNo, Transpose::kNo, + m, n, k, + alpha, + b_buffer, b_offset, b_ld, + temp_symm, 0, k, + beta, + c_buffer, c_offset, c_ld); + + // A and B are now reversed, so also reverse the error codes returned from the Xgemm routine + switch(status) { + case StatusCode::kInvalidMatrixA: status = StatusCode::kInvalidMatrixB; break; + case StatusCode::kInvalidMatrixB: status = StatusCode::kInvalidMatrixA; break; + case StatusCode::kInvalidLeadDimA: status = StatusCode::kInvalidLeadDimB; break; + case StatusCode::kInvalidLeadDimB: status = StatusCode::kInvalidLeadDimA; break; + case StatusCode::kInsufficientMemoryA: status = StatusCode::kInsufficientMemoryB; break; + case StatusCode::kInsufficientMemoryB: status = StatusCode::kInsufficientMemoryA; break; + } + } + + // Return the status of the Xgemm routine + return status; + } catch (...) { return StatusCode::kInvalidKernel; } + } catch (...) { return StatusCode::kTempBufferAllocFailure; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xsymm; +template class Xsymm; +template class Xsymm; +template class Xsymm; +template class Xsymm; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level3/xsyr2k.cc b/src/routines/level3/xsyr2k.cc deleted file mode 100644 index 424d4d2d..00000000 --- a/src/routines/level3/xsyr2k.cc +++ /dev/null @@ -1,210 +0,0 @@ - -// ================================================================================================= -// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This -// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max- -// width of 100 characters per line. -// -// Author(s): -// Cedric Nugteren -// -// This file implements the Xsyr2k class (see the header for information about the class). -// -// ================================================================================================= - -#include "routines/level3/xsyr2k.hpp" - -#include -#include - -namespace clblast { -// ================================================================================================= - -// Constructor: forwards to base class constructor -template -Xsyr2k::Xsyr2k(Queue &queue, EventPointer event, const std::string &name): - Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue()) { - source_string_ = - #include "../../kernels/level3/level3.opencl" - #include "../../kernels/level3/copy_fast.opencl" - #include "../../kernels/level3/copy_pad.opencl" - #include "../../kernels/level3/transpose_fast.opencl" - #include "../../kernels/level3/transpose_pad.opencl" - #include "../../kernels/level3/xgemm_part1.opencl" - #include "../../kernels/level3/xgemm_part2.opencl" - ; -} - -// ================================================================================================= - -// The main routine -template -StatusCode Xsyr2k::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 T beta, - const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { - - // Makes sure all dimensions are larger than zero - if ((n == 0) || (k == 0) ) { return StatusCode::kInvalidDimension; } - - // Computes whether or not the matrices are transposed in memory. This is based on their layout - // (row or column-major) and whether or not they are requested to be pre-transposed. - auto ab_rotated = (layout == Layout::kColMajor && ab_transpose != Transpose::kNo) || - (layout == Layout::kRowMajor && ab_transpose == Transpose::kNo); - auto c_rotated = (layout == Layout::kRowMajor); - - // Computes the first and second dimensions of the A and B matrices taking the layout into account - auto ab_one = (ab_rotated) ? k : n; - auto ab_two = (ab_rotated) ? n : k; - - // Tests the matrices (A, B, C) for validity, first from a perspective of the OpenCL buffers and - // their sizes, and then from a perspective of parameter values (e.g. n, k). Tests whether the - // OpenCL buffers are valid and non-zero and whether the OpenCL buffers have sufficient storage - // space. Also tests that the leading dimensions of: - // matrix A cannot be less than N when rotated, or less than K when not-rotated - // matrix B cannot be less than N when rotated, or less than K when not-rotated - // matrix C cannot be less than N - auto status = TestMatrixA(ab_one, ab_two, a_buffer, a_offset, a_ld); - if (ErrorIn(status)) { return status; } - status = TestMatrixB(ab_one, ab_two, b_buffer, b_offset, b_ld); - if (ErrorIn(status)) { return status; } - status = TestMatrixC(n, n, c_buffer, c_offset, c_ld); - if (ErrorIn(status)) { return status; } - - // Calculates the ceiled versions of n and k - auto n_ceiled = Ceil(n, db_["NWG"]); - auto k_ceiled = Ceil(k, db_["KWG"]); - - // Decides which kernel to run: the upper-triangular or lower-triangular version - auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower"; - - // The padded/transposed input/output matrices: if memory allocation fails, throw an exception - try { - - // Loads the program from the database - const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); - - // Determines whether or not temporary matrices are needed - auto a_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && - ab_rotated == false; - auto b_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && - ab_rotated == false; - - // Creates the temporary matrices - auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto b_temp = (b_no_temp) ? b_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto c_temp = Buffer(context_, n_ceiled*n_ceiled); - - // Upload the scalar arguments as constant buffers to the device (needed for half-precision) - auto alpha_buffer = Buffer(context_, 1); - auto beta_buffer = Buffer(context_, 1); - alpha_buffer.Write(queue_, 1, &alpha); - beta_buffer.Write(queue_, 1, &beta); - - // Events of all kernels (including pre/post processing kernels) - auto eventWaitList = std::vector(); - auto emptyEventList = std::vector(); - - // Runs the pre-processing kernels. This transposes the matrices A and B, but also pads zeros to - // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In - // case nothing has to be done, these kernels can be skipped. - if (!a_no_temp) { - auto eventProcessA = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList, - ab_one, ab_two, a_ld, a_offset, a_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, a_temp, - ConstantOne(), program, - true, ab_rotated, false); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventProcessA); - } - if (!b_no_temp) { - auto eventProcessB = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB.pointer(), emptyEventList, - ab_one, ab_two, b_ld, b_offset, b_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, b_temp, - ConstantOne(), program, - true, ab_rotated, false); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventProcessB); - } - - // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to - // modify the other triangle. - auto eventProcessC = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, - n, n, c_ld, c_offset, c_buffer, - n_ceiled, n_ceiled, n_ceiled, 0, c_temp, - ConstantOne(), program, - true, c_rotated, false); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventProcessC); - - // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary - try { - auto kernel = Kernel(program, kernel_name); - - // Sets the kernel arguments - kernel.SetArgument(0, static_cast(n_ceiled)); - kernel.SetArgument(1, static_cast(k_ceiled)); - kernel.SetArgument(2, alpha_buffer()); - kernel.SetArgument(3, beta_buffer()); - kernel.SetArgument(4, a_temp()); - kernel.SetArgument(5, b_temp()); - kernel.SetArgument(6, c_temp()); - - // Computes the global and local thread sizes - auto global = std::vector{ - (n_ceiled * db_["MDIMC"]) / db_["MWG"], - (n_ceiled * db_["NDIMC"]) / db_["NWG"] - }; - auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; - - // Launches the kernel - auto eventKernel1 = Event(); - status = RunKernel(kernel, queue_, device_, global, local, eventKernel1.pointer(), eventWaitList); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventKernel1); - - // Swaps the arguments for matrices A and B, and sets 'beta' to 1 - auto one = static_cast(1); - beta_buffer.Write(queue_, 1, &one); - kernel.SetArgument(3, beta_buffer()); - kernel.SetArgument(4, b_temp()); - kernel.SetArgument(5, a_temp()); - - // Runs the kernel again - auto eventKernel2 = Event(); - status = RunKernel(kernel, queue_, device_, global, local, eventKernel2.pointer(), eventWaitList); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventKernel2); - - // Runs the post-processing kernel - auto upper = (triangle == Triangle::kUpper); - auto lower = (triangle == Triangle::kLower); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, - n_ceiled, n_ceiled, n_ceiled, 0, c_temp, - n, n, c_ld, c_offset, c_buffer, - ConstantOne(), program, - false, c_rotated, false, upper, lower, false); - if (ErrorIn(status)) { return status; } - - // Successfully finished the computation - return StatusCode::kSuccess; - } catch (...) { return StatusCode::kInvalidKernel; } - } catch (...) { return StatusCode::kTempBufferAllocFailure; } -} - -// ================================================================================================= - -// Compiles the templated class -template class Xsyr2k; -template class Xsyr2k; -template class Xsyr2k; -template class Xsyr2k; -template class Xsyr2k; - -// ================================================================================================= -} // namespace clblast diff --git a/src/routines/level3/xsyr2k.cpp b/src/routines/level3/xsyr2k.cpp new file mode 100644 index 00000000..424d4d2d --- /dev/null +++ b/src/routines/level3/xsyr2k.cpp @@ -0,0 +1,210 @@ + +// ================================================================================================= +// 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 Xsyr2k class (see the header for information about the class). +// +// ================================================================================================= + +#include "routines/level3/xsyr2k.hpp" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xsyr2k::Xsyr2k(Queue &queue, EventPointer event, const std::string &name): + Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue()) { + source_string_ = + #include "../../kernels/level3/level3.opencl" + #include "../../kernels/level3/copy_fast.opencl" + #include "../../kernels/level3/copy_pad.opencl" + #include "../../kernels/level3/transpose_fast.opencl" + #include "../../kernels/level3/transpose_pad.opencl" + #include "../../kernels/level3/xgemm_part1.opencl" + #include "../../kernels/level3/xgemm_part2.opencl" + ; +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xsyr2k::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 T beta, + const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { + + // Makes sure all dimensions are larger than zero + if ((n == 0) || (k == 0) ) { return StatusCode::kInvalidDimension; } + + // Computes whether or not the matrices are transposed in memory. This is based on their layout + // (row or column-major) and whether or not they are requested to be pre-transposed. + auto ab_rotated = (layout == Layout::kColMajor && ab_transpose != Transpose::kNo) || + (layout == Layout::kRowMajor && ab_transpose == Transpose::kNo); + auto c_rotated = (layout == Layout::kRowMajor); + + // Computes the first and second dimensions of the A and B matrices taking the layout into account + auto ab_one = (ab_rotated) ? k : n; + auto ab_two = (ab_rotated) ? n : k; + + // Tests the matrices (A, B, C) for validity, first from a perspective of the OpenCL buffers and + // their sizes, and then from a perspective of parameter values (e.g. n, k). Tests whether the + // OpenCL buffers are valid and non-zero and whether the OpenCL buffers have sufficient storage + // space. Also tests that the leading dimensions of: + // matrix A cannot be less than N when rotated, or less than K when not-rotated + // matrix B cannot be less than N when rotated, or less than K when not-rotated + // matrix C cannot be less than N + auto status = TestMatrixA(ab_one, ab_two, a_buffer, a_offset, a_ld); + if (ErrorIn(status)) { return status; } + status = TestMatrixB(ab_one, ab_two, b_buffer, b_offset, b_ld); + if (ErrorIn(status)) { return status; } + status = TestMatrixC(n, n, c_buffer, c_offset, c_ld); + if (ErrorIn(status)) { return status; } + + // Calculates the ceiled versions of n and k + auto n_ceiled = Ceil(n, db_["NWG"]); + auto k_ceiled = Ceil(k, db_["KWG"]); + + // Decides which kernel to run: the upper-triangular or lower-triangular version + auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower"; + + // The padded/transposed input/output matrices: if memory allocation fails, throw an exception + try { + + // Loads the program from the database + const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); + + // Determines whether or not temporary matrices are needed + auto a_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && + ab_rotated == false; + auto b_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && + ab_rotated == false; + + // Creates the temporary matrices + auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); + auto b_temp = (b_no_temp) ? b_buffer : Buffer(context_, k_ceiled*n_ceiled); + auto c_temp = Buffer(context_, n_ceiled*n_ceiled); + + // Upload the scalar arguments as constant buffers to the device (needed for half-precision) + auto alpha_buffer = Buffer(context_, 1); + auto beta_buffer = Buffer(context_, 1); + alpha_buffer.Write(queue_, 1, &alpha); + beta_buffer.Write(queue_, 1, &beta); + + // Events of all kernels (including pre/post processing kernels) + auto eventWaitList = std::vector(); + auto emptyEventList = std::vector(); + + // Runs the pre-processing kernels. This transposes the matrices A and B, but also pads zeros to + // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In + // case nothing has to be done, these kernels can be skipped. + if (!a_no_temp) { + auto eventProcessA = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList, + ab_one, ab_two, a_ld, a_offset, a_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, a_temp, + ConstantOne(), program, + true, ab_rotated, false); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessA); + } + if (!b_no_temp) { + auto eventProcessB = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessB.pointer(), emptyEventList, + ab_one, ab_two, b_ld, b_offset, b_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, b_temp, + ConstantOne(), program, + true, ab_rotated, false); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessB); + } + + // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to + // modify the other triangle. + auto eventProcessC = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, + n, n, c_ld, c_offset, c_buffer, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + ConstantOne(), program, + true, c_rotated, false); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessC); + + // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary + try { + auto kernel = Kernel(program, kernel_name); + + // Sets the kernel arguments + kernel.SetArgument(0, static_cast(n_ceiled)); + kernel.SetArgument(1, static_cast(k_ceiled)); + kernel.SetArgument(2, alpha_buffer()); + kernel.SetArgument(3, beta_buffer()); + kernel.SetArgument(4, a_temp()); + kernel.SetArgument(5, b_temp()); + kernel.SetArgument(6, c_temp()); + + // Computes the global and local thread sizes + auto global = std::vector{ + (n_ceiled * db_["MDIMC"]) / db_["MWG"], + (n_ceiled * db_["NDIMC"]) / db_["NWG"] + }; + auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; + + // Launches the kernel + auto eventKernel1 = Event(); + status = RunKernel(kernel, queue_, device_, global, local, eventKernel1.pointer(), eventWaitList); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventKernel1); + + // Swaps the arguments for matrices A and B, and sets 'beta' to 1 + auto one = static_cast(1); + beta_buffer.Write(queue_, 1, &one); + kernel.SetArgument(3, beta_buffer()); + kernel.SetArgument(4, b_temp()); + kernel.SetArgument(5, a_temp()); + + // Runs the kernel again + auto eventKernel2 = Event(); + status = RunKernel(kernel, queue_, device_, global, local, eventKernel2.pointer(), eventWaitList); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventKernel2); + + // Runs the post-processing kernel + auto upper = (triangle == Triangle::kUpper); + auto lower = (triangle == Triangle::kLower); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + n, n, c_ld, c_offset, c_buffer, + ConstantOne(), program, + false, c_rotated, false, upper, lower, false); + if (ErrorIn(status)) { return status; } + + // Successfully finished the computation + return StatusCode::kSuccess; + } catch (...) { return StatusCode::kInvalidKernel; } + } catch (...) { return StatusCode::kTempBufferAllocFailure; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xsyr2k; +template class Xsyr2k; +template class Xsyr2k; +template class Xsyr2k; +template class Xsyr2k; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level3/xsyrk.cc b/src/routines/level3/xsyrk.cc deleted file mode 100644 index f56c232b..00000000 --- a/src/routines/level3/xsyrk.cc +++ /dev/null @@ -1,181 +0,0 @@ - -// ================================================================================================= -// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This -// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max- -// width of 100 characters per line. -// -// Author(s): -// Cedric Nugteren -// -// This file implements the Xsyrk class (see the header for information about the class). -// -// ================================================================================================= - -#include "routines/level3/xsyrk.hpp" - -#include -#include - -namespace clblast { -// ================================================================================================= - -// Constructor: forwards to base class constructor -template -Xsyrk::Xsyrk(Queue &queue, EventPointer event, const std::string &name): - Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue()) { - source_string_ = - #include "../../kernels/level3/level3.opencl" - #include "../../kernels/level3/copy_fast.opencl" - #include "../../kernels/level3/copy_pad.opencl" - #include "../../kernels/level3/transpose_fast.opencl" - #include "../../kernels/level3/transpose_pad.opencl" - #include "../../kernels/level3/xgemm_part1.opencl" - #include "../../kernels/level3/xgemm_part2.opencl" - ; -} - -// ================================================================================================= - -// The main routine -template -StatusCode Xsyrk::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 T beta, - const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { - - // Makes sure all dimensions are larger than zero - if ((n == 0) || (k == 0) ) { return StatusCode::kInvalidDimension; } - - // Computes whether or not the matrices are transposed in memory. This is based on their layout - // (row or column-major) and whether or not they are requested to be pre-transposed. - auto a_rotated = (layout == Layout::kColMajor && a_transpose != Transpose::kNo) || - (layout == Layout::kRowMajor && a_transpose == Transpose::kNo); - auto c_rotated = (layout == Layout::kRowMajor); - - // Computes the first and second dimensions of the A matrix taking the layout into account - auto a_one = (a_rotated) ? k : n; - auto a_two = (a_rotated) ? n : k; - - // Tests the two matrices (A, C) for validity, first from a perspective of the OpenCL buffers and - // their sizes, and then from a perspective of parameter values (e.g. n, k). Tests whether the - // OpenCL buffers are valid and non-zero and whether the OpenCL buffers have sufficient storage - // space. Also tests that the leading dimensions of: - // matrix A cannot be less than N when rotated, or less than K when not-rotated - // matrix C cannot be less than N - auto status = TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld); - if (ErrorIn(status)) { return status; } - status = TestMatrixC(n, n, c_buffer, c_offset, c_ld); - if (ErrorIn(status)) { return status; } - - // Calculates the ceiled versions of n and k - auto n_ceiled = Ceil(n, db_["NWG"]); - auto k_ceiled = Ceil(k, db_["KWG"]); - - // Decides which kernel to run: the upper-triangular or lower-triangular version - auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower"; - - // The padded/transposed input/output matrices: if memory allocation fails, throw an exception - try { - - // Loads the program from the database - const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); - - // Determines whether or not temporary matrices are needed - auto a_no_temp = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && - a_rotated == false; - - // Creates the temporary matrices - auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto c_temp = Buffer(context_, n_ceiled*n_ceiled); - - // Upload the scalar arguments as constant buffers to the device (needed for half-precision) - auto alpha_buffer = Buffer(context_, 1); - auto beta_buffer = Buffer(context_, 1); - alpha_buffer.Write(queue_, 1, &alpha); - beta_buffer.Write(queue_, 1, &beta); - - // Events of all kernels (including pre/post processing kernels) - auto eventWaitList = std::vector(); - auto emptyEventList = std::vector(); - - // Runs the pre-processing kernel for matrix A. This transposes the matrix, but also pads zeros - // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In - // case nothing has to be done, these kernels can be skipped. - if (!a_no_temp) { - auto eventProcessA = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList, - a_one, a_two, a_ld, a_offset, a_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, a_temp, - ConstantOne(), program, - true, a_rotated, false); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventProcessA); - } - - // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to - // modify the other triangle. - auto eventProcessC = Event(); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, - n, n, c_ld, c_offset, c_buffer, - n_ceiled, n_ceiled, n_ceiled, 0, c_temp, - ConstantOne(), program, - true, c_rotated, false); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventProcessC); - - // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary - try { - auto kernel = Kernel(program, kernel_name); - - // Sets the kernel arguments - kernel.SetArgument(0, static_cast(n_ceiled)); - kernel.SetArgument(1, static_cast(k_ceiled)); - kernel.SetArgument(2, alpha_buffer()); - kernel.SetArgument(3, beta_buffer()); - kernel.SetArgument(4, a_temp()); - kernel.SetArgument(5, a_temp()); - kernel.SetArgument(6, c_temp()); - - // Computes the global and local thread sizes - auto global = std::vector{ - (n_ceiled * db_["MDIMC"]) / db_["MWG"], - (n_ceiled * db_["NDIMC"]) / db_["NWG"] - }; - auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; - - // Launches the kernel - auto eventKernel = Event(); - status = RunKernel(kernel, queue_, device_, global, local, eventKernel.pointer(), eventWaitList); - if (ErrorIn(status)) { return status; } - eventWaitList.push_back(eventKernel); - - // Runs the post-processing kernel - auto upper = (triangle == Triangle::kUpper); - auto lower = (triangle == Triangle::kLower); - status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, - n_ceiled, n_ceiled, n_ceiled, 0, c_temp, - n, n, c_ld, c_offset, c_buffer, - ConstantOne(), program, - false, c_rotated, false, upper, lower, false); - if (ErrorIn(status)) { return status; } - - - // Successfully finished the computation - return StatusCode::kSuccess; - } catch (...) { return StatusCode::kInvalidKernel; } - } catch (...) { return StatusCode::kTempBufferAllocFailure; } -} - -// ================================================================================================= - -// Compiles the templated class -template class Xsyrk; -template class Xsyrk; -template class Xsyrk; -template class Xsyrk; -template class Xsyrk; - -// ================================================================================================= -} // namespace clblast diff --git a/src/routines/level3/xsyrk.cpp b/src/routines/level3/xsyrk.cpp new file mode 100644 index 00000000..f56c232b --- /dev/null +++ b/src/routines/level3/xsyrk.cpp @@ -0,0 +1,181 @@ + +// ================================================================================================= +// 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 Xsyrk class (see the header for information about the class). +// +// ================================================================================================= + +#include "routines/level3/xsyrk.hpp" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xsyrk::Xsyrk(Queue &queue, EventPointer event, const std::string &name): + Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue()) { + source_string_ = + #include "../../kernels/level3/level3.opencl" + #include "../../kernels/level3/copy_fast.opencl" + #include "../../kernels/level3/copy_pad.opencl" + #include "../../kernels/level3/transpose_fast.opencl" + #include "../../kernels/level3/transpose_pad.opencl" + #include "../../kernels/level3/xgemm_part1.opencl" + #include "../../kernels/level3/xgemm_part2.opencl" + ; +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xsyrk::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 T beta, + const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { + + // Makes sure all dimensions are larger than zero + if ((n == 0) || (k == 0) ) { return StatusCode::kInvalidDimension; } + + // Computes whether or not the matrices are transposed in memory. This is based on their layout + // (row or column-major) and whether or not they are requested to be pre-transposed. + auto a_rotated = (layout == Layout::kColMajor && a_transpose != Transpose::kNo) || + (layout == Layout::kRowMajor && a_transpose == Transpose::kNo); + auto c_rotated = (layout == Layout::kRowMajor); + + // Computes the first and second dimensions of the A matrix taking the layout into account + auto a_one = (a_rotated) ? k : n; + auto a_two = (a_rotated) ? n : k; + + // Tests the two matrices (A, C) for validity, first from a perspective of the OpenCL buffers and + // their sizes, and then from a perspective of parameter values (e.g. n, k). Tests whether the + // OpenCL buffers are valid and non-zero and whether the OpenCL buffers have sufficient storage + // space. Also tests that the leading dimensions of: + // matrix A cannot be less than N when rotated, or less than K when not-rotated + // matrix C cannot be less than N + auto status = TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld); + if (ErrorIn(status)) { return status; } + status = TestMatrixC(n, n, c_buffer, c_offset, c_ld); + if (ErrorIn(status)) { return status; } + + // Calculates the ceiled versions of n and k + auto n_ceiled = Ceil(n, db_["NWG"]); + auto k_ceiled = Ceil(k, db_["KWG"]); + + // Decides which kernel to run: the upper-triangular or lower-triangular version + auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower"; + + // The padded/transposed input/output matrices: if memory allocation fails, throw an exception + try { + + // Loads the program from the database + const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); + + // Determines whether or not temporary matrices are needed + auto a_no_temp = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && + a_rotated == false; + + // Creates the temporary matrices + auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); + auto c_temp = Buffer(context_, n_ceiled*n_ceiled); + + // Upload the scalar arguments as constant buffers to the device (needed for half-precision) + auto alpha_buffer = Buffer(context_, 1); + auto beta_buffer = Buffer(context_, 1); + alpha_buffer.Write(queue_, 1, &alpha); + beta_buffer.Write(queue_, 1, &beta); + + // Events of all kernels (including pre/post processing kernels) + auto eventWaitList = std::vector(); + auto emptyEventList = std::vector(); + + // Runs the pre-processing kernel for matrix A. This transposes the matrix, but also pads zeros + // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In + // case nothing has to be done, these kernels can be skipped. + if (!a_no_temp) { + auto eventProcessA = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessA.pointer(), emptyEventList, + a_one, a_two, a_ld, a_offset, a_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, a_temp, + ConstantOne(), program, + true, a_rotated, false); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessA); + } + + // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to + // modify the other triangle. + auto eventProcessC = Event(); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, eventProcessC.pointer(), emptyEventList, + n, n, c_ld, c_offset, c_buffer, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + ConstantOne(), program, + true, c_rotated, false); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventProcessC); + + // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary + try { + auto kernel = Kernel(program, kernel_name); + + // Sets the kernel arguments + kernel.SetArgument(0, static_cast(n_ceiled)); + kernel.SetArgument(1, static_cast(k_ceiled)); + kernel.SetArgument(2, alpha_buffer()); + kernel.SetArgument(3, beta_buffer()); + kernel.SetArgument(4, a_temp()); + kernel.SetArgument(5, a_temp()); + kernel.SetArgument(6, c_temp()); + + // Computes the global and local thread sizes + auto global = std::vector{ + (n_ceiled * db_["MDIMC"]) / db_["MWG"], + (n_ceiled * db_["NDIMC"]) / db_["NWG"] + }; + auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; + + // Launches the kernel + auto eventKernel = Event(); + status = RunKernel(kernel, queue_, device_, global, local, eventKernel.pointer(), eventWaitList); + if (ErrorIn(status)) { return status; } + eventWaitList.push_back(eventKernel); + + // Runs the post-processing kernel + auto upper = (triangle == Triangle::kUpper); + auto lower = (triangle == Triangle::kLower); + status = PadCopyTransposeMatrix(queue_, device_, context_, db_, event_, eventWaitList, + n_ceiled, n_ceiled, n_ceiled, 0, c_temp, + n, n, c_ld, c_offset, c_buffer, + ConstantOne(), program, + false, c_rotated, false, upper, lower, false); + if (ErrorIn(status)) { return status; } + + + // Successfully finished the computation + return StatusCode::kSuccess; + } catch (...) { return StatusCode::kInvalidKernel; } + } catch (...) { return StatusCode::kTempBufferAllocFailure; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xsyrk; +template class Xsyrk; +template class Xsyrk; +template class Xsyrk; +template class Xsyrk; + +// ================================================================================================= +} // namespace clblast diff --git a/src/routines/level3/xtrmm.cc b/src/routines/level3/xtrmm.cc deleted file mode 100644 index 74a82822..00000000 --- a/src/routines/level3/xtrmm.cc +++ /dev/null @@ -1,140 +0,0 @@ - -// ================================================================================================= -// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This -// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max- -// width of 100 characters per line. -// -// Author(s): -// Cedric Nugteren -// -// This file implements the Xtrmm class (see the header for information about the class). -// -// ================================================================================================= - -#include "routines/level3/xtrmm.hpp" - -#include -#include - -namespace clblast { -// ================================================================================================= - -// Constructor: forwards to base class constructor -template -Xtrmm::Xtrmm(Queue &queue, EventPointer event, const std::string &name): - Xgemm(queue, event, name) { -} - -// ================================================================================================= - -// The main routine -template -StatusCode Xtrmm::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) { - - // Makes sure all dimensions are larger than zero - if ((m == 0) || (n == 0)) { return StatusCode::kInvalidDimension; } - - // Computes the k dimension. This is based on whether or not matrix is A (on the left) - // or B (on the right) in the Xgemm routine. - auto k = (side == Side::kLeft) ? m : n; - - // Checks for validity of the triangular A matrix - auto status = TestMatrixA(k, k, a_buffer, a_offset, a_ld); - if (ErrorIn(status)) { return status; } - - // Determines which kernel to run based on the layout (the Xgemm kernel assumes column-major as - // default) and on whether we are dealing with an upper or lower triangle of the triangular matrix - bool is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) || - (triangle == Triangle::kLower && layout == Layout::kRowMajor)); - auto kernel_name = (is_upper) ? "TriaUpperToSquared" : "TriaLowerToSquared"; - - // Determines whether or not the triangular matrix is unit-diagonal - auto unit_diagonal = (diagonal == Diagonal::kUnit) ? true : false; - - // Temporary buffer for a copy of the triangular matrix - try { - auto temp_triangular = Buffer(context_, k*k); - - // Creates a general matrix from the triangular matrix to be able to run the regular Xgemm - // routine afterwards - try { - const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); - auto kernel = Kernel(program, kernel_name); - - // Sets the arguments for the triangular-to-squared kernel - kernel.SetArgument(0, static_cast(k)); - kernel.SetArgument(1, static_cast(a_ld)); - kernel.SetArgument(2, static_cast(a_offset)); - kernel.SetArgument(3, a_buffer()); - kernel.SetArgument(4, static_cast(k)); - kernel.SetArgument(5, static_cast(k)); - kernel.SetArgument(6, static_cast(0)); - kernel.SetArgument(7, temp_triangular()); - kernel.SetArgument(8, static_cast(unit_diagonal)); - - // Uses the common padding kernel's thread configuration. This is allowed, since the - // triangular-to-squared kernel uses the same parameters. - auto global = std::vector{Ceil(CeilDiv(k, db_["PAD_WPTX"]), db_["PAD_DIMX"]), - Ceil(CeilDiv(k, db_["PAD_WPTY"]), db_["PAD_DIMY"])}; - auto local = std::vector{db_["PAD_DIMX"], db_["PAD_DIMY"]}; - auto kernelEvent = Event(); - status = RunKernel(kernel, queue_, device_, global, local, kernelEvent.pointer()); - if (ErrorIn(status)) { return status; } - - // Synchronize now: 'DoGemm' does not accept a list of events to wait for - kernelEvent.WaitForCompletion(); - - // Runs the regular Xgemm code with either "B := alpha*A*B" or ... - if (side == Side::kLeft) { - status = DoGemm(layout, a_transpose, Transpose::kNo, - m, n, k, - alpha, - temp_triangular, 0, k, - b_buffer, b_offset, b_ld, - static_cast(0.0), - b_buffer, b_offset, b_ld); - } - - // ... with "B := alpha*B*A". Note that A and B are now reversed. - else { - status = DoGemm(layout, Transpose::kNo, a_transpose, - m, n, k, - alpha, - b_buffer, b_offset, b_ld, - temp_triangular, 0, k, - static_cast(0.0), - b_buffer, b_offset, b_ld); - - // A and B are now reversed, so also reverse the error codes returned from the Xgemm routine - switch(status) { - case StatusCode::kInvalidMatrixA: status = StatusCode::kInvalidMatrixB; break; - case StatusCode::kInvalidMatrixB: status = StatusCode::kInvalidMatrixA; break; - case StatusCode::kInvalidLeadDimA: status = StatusCode::kInvalidLeadDimB; break; - case StatusCode::kInvalidLeadDimB: status = StatusCode::kInvalidLeadDimA; break; - case StatusCode::kInsufficientMemoryA: status = StatusCode::kInsufficientMemoryB; break; - case StatusCode::kInsufficientMemoryB: status = StatusCode::kInsufficientMemoryA; break; - } - } - - // Return the status of the Xgemm routine - return status; - } catch (...) { return StatusCode::kInvalidKernel; } - } catch (...) { return StatusCode::kTempBufferAllocFailure; } -} - -// ================================================================================================= - -// Compiles the templated class -template class Xtrmm; -template class Xtrmm; -template class Xtrmm; -template class Xtrmm; -template class Xtrmm; - -// ================================================================================================= -} // namespace clblast diff --git a/src/routines/level3/xtrmm.cpp b/src/routines/level3/xtrmm.cpp new file mode 100644 index 00000000..74a82822 --- /dev/null +++ b/src/routines/level3/xtrmm.cpp @@ -0,0 +1,140 @@ + +// ================================================================================================= +// 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 Xtrmm class (see the header for information about the class). +// +// ================================================================================================= + +#include "routines/level3/xtrmm.hpp" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xtrmm::Xtrmm(Queue &queue, EventPointer event, const std::string &name): + Xgemm(queue, event, name) { +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xtrmm::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) { + + // Makes sure all dimensions are larger than zero + if ((m == 0) || (n == 0)) { return StatusCode::kInvalidDimension; } + + // Computes the k dimension. This is based on whether or not matrix is A (on the left) + // or B (on the right) in the Xgemm routine. + auto k = (side == Side::kLeft) ? m : n; + + // Checks for validity of the triangular A matrix + auto status = TestMatrixA(k, k, a_buffer, a_offset, a_ld); + if (ErrorIn(status)) { return status; } + + // Determines which kernel to run based on the layout (the Xgemm kernel assumes column-major as + // default) and on whether we are dealing with an upper or lower triangle of the triangular matrix + bool is_upper = ((triangle == Triangle::kUpper && layout != Layout::kRowMajor) || + (triangle == Triangle::kLower && layout == Layout::kRowMajor)); + auto kernel_name = (is_upper) ? "TriaUpperToSquared" : "TriaLowerToSquared"; + + // Determines whether or not the triangular matrix is unit-diagonal + auto unit_diagonal = (diagonal == Diagonal::kUnit) ? true : false; + + // Temporary buffer for a copy of the triangular matrix + try { + auto temp_triangular = Buffer(context_, k*k); + + // Creates a general matrix from the triangular matrix to be able to run the regular Xgemm + // routine afterwards + try { + const auto program = GetProgramFromCache(context_, PrecisionValue(), routine_name_); + auto kernel = Kernel(program, kernel_name); + + // Sets the arguments for the triangular-to-squared kernel + kernel.SetArgument(0, static_cast(k)); + kernel.SetArgument(1, static_cast(a_ld)); + kernel.SetArgument(2, static_cast(a_offset)); + kernel.SetArgument(3, a_buffer()); + kernel.SetArgument(4, static_cast(k)); + kernel.SetArgument(5, static_cast(k)); + kernel.SetArgument(6, static_cast(0)); + kernel.SetArgument(7, temp_triangular()); + kernel.SetArgument(8, static_cast(unit_diagonal)); + + // Uses the common padding kernel's thread configuration. This is allowed, since the + // triangular-to-squared kernel uses the same parameters. + auto global = std::vector{Ceil(CeilDiv(k, db_["PAD_WPTX"]), db_["PAD_DIMX"]), + Ceil(CeilDiv(k, db_["PAD_WPTY"]), db_["PAD_DIMY"])}; + auto local = std::vector{db_["PAD_DIMX"], db_["PAD_DIMY"]}; + auto kernelEvent = Event(); + status = RunKernel(kernel, queue_, device_, global, local, kernelEvent.pointer()); + if (ErrorIn(status)) { return status; } + + // Synchronize now: 'DoGemm' does not accept a list of events to wait for + kernelEvent.WaitForCompletion(); + + // Runs the regular Xgemm code with either "B := alpha*A*B" or ... + if (side == Side::kLeft) { + status = DoGemm(layout, a_transpose, Transpose::kNo, + m, n, k, + alpha, + temp_triangular, 0, k, + b_buffer, b_offset, b_ld, + static_cast(0.0), + b_buffer, b_offset, b_ld); + } + + // ... with "B := alpha*B*A". Note that A and B are now reversed. + else { + status = DoGemm(layout, Transpose::kNo, a_transpose, + m, n, k, + alpha, + b_buffer, b_offset, b_ld, + temp_triangular, 0, k, + static_cast(0.0), + b_buffer, b_offset, b_ld); + + // A and B are now reversed, so also reverse the error codes returned from the Xgemm routine + switch(status) { + case StatusCode::kInvalidMatrixA: status = StatusCode::kInvalidMatrixB; break; + case StatusCode::kInvalidMatrixB: status = StatusCode::kInvalidMatrixA; break; + case StatusCode::kInvalidLeadDimA: status = StatusCode::kInvalidLeadDimB; break; + case StatusCode::kInvalidLeadDimB: status = StatusCode::kInvalidLeadDimA; break; + case StatusCode::kInsufficientMemoryA: status = StatusCode::kInsufficientMemoryB; break; + case StatusCode::kInsufficientMemoryB: status = StatusCode::kInsufficientMemoryA; break; + } + } + + // Return the status of the Xgemm routine + return status; + } catch (...) { return StatusCode::kInvalidKernel; } + } catch (...) { return StatusCode::kTempBufferAllocFailure; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xtrmm; +template class Xtrmm; +template class Xtrmm; +template class Xtrmm; +template class Xtrmm; + +// ================================================================================================= +} // namespace clblast -- cgit v1.2.3