diff options
author | Cedric Nugteren <web@cedricnugteren.nl> | 2016-06-19 13:55:49 +0200 |
---|---|---|
committer | Cedric Nugteren <web@cedricnugteren.nl> | 2016-06-19 13:55:49 +0200 |
commit | 61203453aaca4e47c05c598a673150522160ca87 (patch) | |
tree | 36cdce32423c2227de88e019c34415bef9499a6a /src/routines/level3/xsyrk.cpp | |
parent | f726fbdc9fef937fbe32222f0e66aac8d7e2678c (diff) |
Renamed all C++ source files to .cpp to match the .hpp extension better
Diffstat (limited to 'src/routines/level3/xsyrk.cpp')
-rw-r--r-- | src/routines/level3/xsyrk.cpp | 181 |
1 files changed, 181 insertions, 0 deletions
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 <www.cedricnugteren.nl> +// +// This file implements the Xsyrk class (see the header for information about the class). +// +// ================================================================================================= + +#include "routines/level3/xsyrk.hpp" + +#include <string> +#include <vector> + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template <typename T> +Xsyrk<T>::Xsyrk(Queue &queue, EventPointer event, const std::string &name): + Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue<T>()) { + 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 <typename T> +StatusCode Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const Transpose a_transpose, + const size_t n, const size_t k, + const T alpha, + const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld, + const T beta, + const Buffer<T> &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<T>(), 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<T>(context_, k_ceiled*n_ceiled); + auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled); + + // Upload the scalar arguments as constant buffers to the device (needed for half-precision) + auto alpha_buffer = Buffer<T>(context_, 1); + auto beta_buffer = Buffer<T>(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<Event>(); + auto emptyEventList = std::vector<Event>(); + + // 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<T>(), 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<T>(), 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<int>(n_ceiled)); + kernel.SetArgument(1, static_cast<int>(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<size_t>{ + (n_ceiled * db_["MDIMC"]) / db_["MWG"], + (n_ceiled * db_["NDIMC"]) / db_["NWG"] + }; + auto local = std::vector<size_t>{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<T>(), 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<half>; +template class Xsyrk<float>; +template class Xsyrk<double>; +template class Xsyrk<float2>; +template class Xsyrk<double2>; + +// ================================================================================================= +} // namespace clblast |