diff options
| -rw-r--r-- | CMakeLists.txt | 2 | ||||
| -rw-r--r-- | include/clblast.h | 10 | ||||
| -rw-r--r-- | include/internal/routines/xherk.h | 47 | ||||
| -rw-r--r-- | src/clblast.cc | 47 | ||||
| -rw-r--r-- | src/routines/xherk.cc | 156 | ||||
| -rw-r--r-- | test/performance/routines/xherk.cc | 40 | ||||
| -rw-r--r-- | test/routines/xherk.h | 121 | ||||
| -rw-r--r-- | test/wrapper_clblas.h | 28 |
8 files changed, 449 insertions, 2 deletions
diff --git a/CMakeLists.txt b/CMakeLists.txt index 60060961..61af6b82 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -98,7 +98,7 @@ set(SAMPLE_PROGRAMS sgemm) set(ROUTINES xaxpy xgemv - xgemm xsymm xsyrk xsyr2k xtrmm) + xgemm xsymm xsyrk xherk xsyr2k xtrmm) # ================================================================================================== diff --git a/include/clblast.h b/include/clblast.h index 5da10810..d7c902d9 100644 --- a/include/clblast.h +++ b/include/clblast.h @@ -140,6 +140,16 @@ StatusCode Syrk(const Layout layout, const Triangle triangle, const Transpose a_ cl_mem c_buffer, const size_t c_offset, const size_t c_ld, cl_command_queue* queue, cl_event* event); +// Templated-precision rank-K update of a hermitian matrix: CHERK/ZHERK +template <typename T> +StatusCode Herk(const Layout layout, const Triangle triangle, const Transpose a_transpose, + const size_t n, const size_t k, + const T alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, + const T beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, + cl_command_queue* queue, cl_event* event); + // Templated-precision rank-2K update of a symmetric matrix: SSYR2K/DSYR2K/CSYR2K/ZSYR2K template <typename T> StatusCode Syr2k(const Layout layout, const Triangle triangle, const Transpose ab_transpose, diff --git a/include/internal/routines/xherk.h b/include/internal/routines/xherk.h new file mode 100644 index 00000000..9b361254 --- /dev/null +++ b/include/internal/routines/xherk.h @@ -0,0 +1,47 @@ + +// ================================================================================================= +// 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 Xherk routine. The precision is implemented using the template argument +// 'T', whereas the alpha/beta arguments are of type 'U'. The implementation is very similar to the +// Xsyrk routine. +// +// ================================================================================================= + +#ifndef CLBLAST_ROUTINES_XHERK_H_ +#define CLBLAST_ROUTINES_XHERK_H_ + +#include "internal/routine.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T, typename U> +class Xherk: public Routine { + public: + Xherk(CommandQueue &queue, Event &event); + + // Templated-precision implementation of the routine + StatusCode 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); + + private: + // Static variable to get the precision + const static Precision precision_; +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_ROUTINES_XHERK_H_ +#endif diff --git a/src/clblast.cc b/src/clblast.cc index 299d0a18..638bc944 100644 --- a/src/clblast.cc +++ b/src/clblast.cc @@ -27,6 +27,7 @@ #include "internal/routines/xgemm.h" #include "internal/routines/xsymm.h" #include "internal/routines/xsyrk.h" +#include "internal/routines/xherk.h" #include "internal/routines/xsyr2k.h" #include "internal/routines/xtrmm.h" @@ -302,6 +303,50 @@ template StatusCode Syrk<double2>(const Layout, const Triangle, const Transpose, // ================================================================================================= +// HERK +template <typename T> +StatusCode Herk(const Layout layout, const Triangle triangle, const Transpose a_transpose, + const size_t n, const size_t k, const T alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, const T beta, + cl_mem c_buffer, const size_t c_offset, const size_t c_ld, + cl_command_queue* queue, cl_event* event) { + auto queue_cpp = CommandQueue(*queue); + auto event_cpp = Event(*event); + auto routine = Xherk<std::complex<T>,T>(queue_cpp, event_cpp); + + // Loads the kernel source-code as an include (C++11 raw string literal) + std::string common_source1 = + #include "kernels/copy.opencl" + std::string common_source2 = + #include "kernels/pad.opencl" + std::string common_source3 = + #include "kernels/transpose.opencl" + std::string common_source4 = + #include "kernels/padtranspose.opencl" + std::string kernel_source = + #include "kernels/xgemm.opencl" + auto status = routine.SetUp(common_source1 + common_source2 + common_source3 + common_source4 + + kernel_source); + if (status != StatusCode::kSuccess) { return status; } + + // Runs the routine + return routine.DoHerk(layout, triangle, a_transpose, n, k, alpha, + Buffer(a_buffer), a_offset, a_ld, beta, + Buffer(c_buffer), c_offset, c_ld); +} +template StatusCode Herk<float>(const Layout, const Triangle, const Transpose, + const size_t, const size_t, const float, + const cl_mem, const size_t, const size_t, const float, + cl_mem, const size_t, const size_t, + cl_command_queue*, cl_event*); +template StatusCode Herk<double>(const Layout, const Triangle, const Transpose, + const size_t, const size_t, const double, + const cl_mem, const size_t, const size_t, const double, + cl_mem, const size_t, const size_t, + cl_command_queue*, cl_event*); + +// ================================================================================================= + // SYR2K template <typename T> StatusCode Syr2k(const Layout layout, const Triangle triangle, const Transpose ab_transpose, @@ -456,7 +501,7 @@ StatusCode Trsm(const Layout layout, const Side side, const Triangle triangle, Buffer(a_buffer), a_offset, a_ld, Buffer(b_buffer), b_offset, b_ld); */ - return StatusCode::kSuccess; + return StatusCode::kNotImplemented; } template StatusCode Trsm<float>(const Layout, const Side, const Triangle, const Transpose, const Diagonal, diff --git a/src/routines/xherk.cc b/src/routines/xherk.cc new file mode 100644 index 00000000..6bc9cd6c --- /dev/null +++ b/src/routines/xherk.cc @@ -0,0 +1,156 @@ + +// ================================================================================================= +// 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 Xherk class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/xherk.h" + +#include <string> +#include <vector> + +namespace clblast { +// ================================================================================================= + +// Specific implementations to get the memory-type based on a template argument +template <> const Precision Xherk<float2,float>::precision_ = Precision::kComplexSingle; +template <> const Precision Xherk<double2,double>::precision_ = Precision::kComplexDouble; + +// ================================================================================================= + +// Constructor: forwards to base class constructor +template <typename T, typename U> +Xherk<T,U>::Xherk(CommandQueue &queue, Event &event): + Routine(queue, event, {"Copy", "Pad", "Transpose", "PadTranspose", "Xgemm"}, precision_) { +} + +// ================================================================================================= + +// The main routine +template <typename T, typename U> +StatusCode Xherk<T,U>::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, sizeof(T)); + if (ErrorIn(status)) { return status; } + status = TestMatrixC(n, n, c_buffer, c_offset, c_ld, sizeof(T)); + 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"; + + // Allocates space on the device for padded and/or transposed input and output matrices. + try { + auto temp_a = Buffer(context_, CL_MEM_READ_WRITE, k_ceiled*n_ceiled*sizeof(T)); + auto temp_b = Buffer(context_, CL_MEM_READ_WRITE, k_ceiled*n_ceiled*sizeof(T)); + auto temp_c = Buffer(context_, CL_MEM_READ_WRITE, n_ceiled*n_ceiled*sizeof(T)); + + // Loads the program from the database + auto& program = GetProgramFromCache(); + + // Runs the pre-processing kernel. This transposes the matrix A, but also pads zeros to + // fill it up until it reaches a certain multiple of size (kernel parameter dependent). It + // creates two copies: + status = PadCopyTransposeMatrix(a_one, a_two, a_ld, a_offset, a_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, temp_a, + a_rotated, a_conjugate, true, false, false, false, program); + if (ErrorIn(status)) { return status; } + status = PadCopyTransposeMatrix(a_one, a_two, a_ld, a_offset, a_buffer, + n_ceiled, k_ceiled, n_ceiled, 0, temp_b, + a_rotated, b_conjugate, true, false, false, false, program); + 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. + status = PadCopyTransposeMatrix(n, n, c_ld, c_offset, c_buffer, + n_ceiled, n_ceiled, n_ceiled, 0, temp_c, + c_rotated, false, true, false, false, false, program); + 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 + auto complex_alpha = T{alpha, static_cast<U>(0.0)}; + auto complex_beta = T{beta, static_cast<U>(0.0)}; + kernel.SetArgument(0, static_cast<int>(n_ceiled)); + kernel.SetArgument(1, static_cast<int>(k_ceiled)); + kernel.SetArgument(2, complex_alpha); + kernel.SetArgument(3, complex_beta); + kernel.SetArgument(4, temp_a()); + kernel.SetArgument(5, temp_b()); + kernel.SetArgument(6, temp_c()); + + // 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 + status = RunKernel(kernel, global, local); + if (ErrorIn(status)) { return status; } + + // Runs the post-processing kernel + auto upper = (triangle == Triangle::kUpper); + auto lower = (triangle == Triangle::kLower); + status = PadCopyTransposeMatrix(n_ceiled, n_ceiled, n_ceiled, 0, temp_c, + n, n, c_ld, c_offset, c_buffer, + c_rotated, false, false, upper, lower, true, program); + 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<float2,float>; +template class Xherk<double2,double>; + +// ================================================================================================= +} // namespace clblast diff --git a/test/performance/routines/xherk.cc b/test/performance/routines/xherk.cc new file mode 100644 index 00000000..ce18152e --- /dev/null +++ b/test/performance/routines/xherk.cc @@ -0,0 +1,40 @@ + +// ================================================================================================= +// 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 Xherk command-line interface performance tester. +// +// ================================================================================================= + +#include "performance/client.h" +#include "routines/xherk.h" + +// ================================================================================================= + +// Shortcuts to the clblast namespace +using float2 = clblast::float2; +using double2 = clblast::double2; + +// Main function (not within the clblast namespace) +int main(int argc, char *argv[]) { + switch(clblast::GetPrecision(argc, argv)) { + case clblast::Precision::kHalf: + throw std::runtime_error("Unsupported precision mode"); + case clblast::Precision::kSingle: + throw std::runtime_error("Unsupported precision mode"); + case clblast::Precision::kDouble: + throw std::runtime_error("Unsupported precision mode"); + case clblast::Precision::kComplexSingle: + clblast::RunClient<clblast::TestXherk<float2,float>, float2, float>(argc, argv); break; + case clblast::Precision::kComplexDouble: + clblast::RunClient<clblast::TestXherk<double2,double>, double2, double>(argc, argv); break; + } + return 0; +} + +// ================================================================================================= diff --git a/test/routines/xherk.h b/test/routines/xherk.h new file mode 100644 index 00000000..780b9b52 --- /dev/null +++ b/test/routines/xherk.h @@ -0,0 +1,121 @@ + +// ================================================================================================= +// 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 a class with static methods to describe the Xherk routine. Examples of +// such 'descriptions' are how to calculate the size a of buffer or how to run the routine. These +// static methods are used by the correctness tester and the performance tester. +// +// ================================================================================================= + +#ifndef CLBLAST_TEST_ROUTINES_XHERK_H_ +#define CLBLAST_TEST_ROUTINES_XHERK_H_ + +#include <vector> +#include <string> + +#include "wrapper_clblas.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T, typename U> +class TestXherk { + public: + + // The list of arguments relevant for this routine + static std::vector<std::string> GetOptions() { + return {kArgN, kArgK, + kArgLayout, kArgTriangle, kArgATransp, + kArgALeadDim, kArgCLeadDim, + kArgAOffset, kArgCOffset, + kArgAlpha, kArgBeta}; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeA(const Arguments<U> &args) { + auto a_rotated = (args.layout == Layout::kColMajor && args.a_transpose != Transpose::kNo) || + (args.layout == Layout::kRowMajor && args.a_transpose == Transpose::kNo); + auto a_two = (a_rotated) ? args.n : args.k; + return a_two * args.a_ld + args.a_offset; + } + static size_t GetSizeC(const Arguments<U> &args) { + return args.n * args.c_ld + args.c_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments<U> &args) { + args.a_size = GetSizeA(args); + args.c_size = GetSizeC(args); + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments<U> &args) { return args.k; } + static size_t DefaultLDB(const Arguments<U> &) { return 1; } // N/A for this routine + static size_t DefaultLDC(const Arguments<U> &args) { return args.n; } + + // Describes how to run the CLBlast routine + static StatusCode RunRoutine(const Arguments<U> &args, const Buffers &buffers, + CommandQueue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = Herk(args.layout, args.triangle, args.a_transpose, + args.n, args.k, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, args.beta, + buffers.c_mat(), args.c_offset, args.c_ld, + &queue_plain, &event); + clWaitForEvents(1, &event); + return status; + } + + // Describes how to run the clBLAS routine (for correctness/performance comparison) + static StatusCode RunReference(const Arguments<U> &args, const Buffers &buffers, + CommandQueue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = clblasXherk(static_cast<clblasOrder>(args.layout), + static_cast<clblasUplo>(args.triangle), + static_cast<clblasTranspose>(args.a_transpose), + args.n, args.k, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, args.beta, + buffers.c_mat(), args.c_offset, args.c_ld, + 1, &queue_plain, 0, nullptr, &event); + clWaitForEvents(1, &event); + return static_cast<StatusCode>(status); + } + + // Describes how to download the results of the computation (more importantly: which buffer) + static std::vector<T> DownloadResult(const Arguments<U> &args, Buffers &buffers, + CommandQueue &queue) { + std::vector<T> result(args.c_size, static_cast<T>(0)); + buffers.c_mat.ReadBuffer(queue, args.c_size*sizeof(T), result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments<U> &args) { return args.n; } + static size_t ResultID2(const Arguments<U> &args) { return args.n; } + static size_t GetResultIndex(const Arguments<U> &args, const size_t id1, const size_t id2) { + return id1*args.c_ld + id2 + args.c_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments<U> &args) { + return args.n * args.n * args.k; + } + static size_t GetBytes(const Arguments<U> &args) { + return (args.n*args.k + args.n*args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XHERK_H_ +#endif diff --git a/test/wrapper_clblas.h b/test/wrapper_clblas.h index a03ade5e..360adfc8 100644 --- a/test/wrapper_clblas.h +++ b/test/wrapper_clblas.h @@ -325,6 +325,34 @@ clblasStatus clblasXsyrk( num_queues, queues, num_wait_events, wait_events, events); } +// This calls {clblasCherk, clblasZherk} with the arguments forwarded. +clblasStatus clblasXherk( + clblasOrder layout, clblasUplo triangle, clblasTranspose a_transpose, + size_t n, size_t k, float alpha, + const cl_mem a_mat, size_t a_offset, size_t a_ld, float beta, + const cl_mem c_mat, size_t c_offset, size_t c_ld, + cl_uint num_queues, cl_command_queue *queues, + cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { + return clblasCherk(layout, triangle, a_transpose, + n, k, alpha, + a_mat, a_offset, a_ld, beta, + c_mat, c_offset, c_ld, + num_queues, queues, num_wait_events, wait_events, events); +} +clblasStatus clblasXherk( + clblasOrder layout, clblasUplo triangle, clblasTranspose a_transpose, + size_t n, size_t k, double alpha, + const cl_mem a_mat, size_t a_offset, size_t a_ld, double beta, + const cl_mem c_mat, size_t c_offset, size_t c_ld, + cl_uint num_queues, cl_command_queue *queues, + cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { + return clblasZherk(layout, triangle, a_transpose, + n, k, alpha, + a_mat, a_offset, a_ld, beta, + c_mat, c_offset, c_ld, + num_queues, queues, num_wait_events, wait_events, events); +} + // This calls {clblasSsyr2k, clblasDsyr2k, clblasCsyr2k, clblasZsyr2k} with the arguments forwarded. clblasStatus clblasXsyr2k( clblasOrder layout, clblasUplo triangle, clblasTranspose ab_transpose, |