From b89517a2e77d855debbaabc45861ba6d8b7d4adc Mon Sep 17 00:00:00 2001 From: CNugteren Date: Fri, 31 Jul 2015 17:13:41 +0200 Subject: Added SYMV routine --- CMakeLists.txt | 2 +- include/clblast.h | 11 +++ include/internal/routines/level2/xsymv.h | 58 +++++++++++++ src/clblast.cc | 42 ++++++++++ src/routines/level2/xgemv.cc | 3 +- src/routines/level2/xsymv.cc | 100 +++++++++++++++++++++++ test/correctness/routines/level2/xsymv.cc | 26 ++++++ test/performance/routines/level2/xsymv.cc | 36 +++++++++ test/routines/level2/xsymv.h | 130 ++++++++++++++++++++++++++++++ test/wrapper_clblas.h | 28 +++++++ 10 files changed, 434 insertions(+), 2 deletions(-) create mode 100644 include/internal/routines/level2/xsymv.h create mode 100644 src/routines/level2/xsymv.cc create mode 100644 test/correctness/routines/level2/xsymv.cc create mode 100644 test/performance/routines/level2/xsymv.cc create mode 100644 test/routines/level2/xsymv.h diff --git a/CMakeLists.txt b/CMakeLists.txt index 89dc0fd9..45429dfa 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -105,7 +105,7 @@ include_directories(${clblast_SOURCE_DIR}/include ${OPENCL_INCLUDE_DIRS}) set(KERNELS copy pad transpose padtranspose xaxpy xgemv xgemm) set(SAMPLE_PROGRAMS sgemm) set(LEVEL1_ROUTINES xaxpy) -set(LEVEL2_ROUTINES xgemv) +set(LEVEL2_ROUTINES xgemv xsymv) set(LEVEL3_ROUTINES xgemm xsymm xhemm xsyrk xherk xsyr2k xher2k xtrmm) set(ROUTINES ${LEVEL1_ROUTINES} ${LEVEL2_ROUTINES} ${LEVEL3_ROUTINES}) diff --git a/include/clblast.h b/include/clblast.h index 80ea1707..93acc03d 100644 --- a/include/clblast.h +++ b/include/clblast.h @@ -105,6 +105,17 @@ StatusCode Gemv(const Layout layout, const Transpose a_transpose, cl_mem y_buffer, const size_t y_offset, const size_t y_inc, cl_command_queue* queue, cl_event* event); +// Templated-precision symmetric matrix-vector multiplication: SSYMV/DSYMV +template +StatusCode Symv(const Layout layout, const Triangle triangle, + const size_t n, + const T alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, + const cl_mem x_buffer, const size_t x_offset, const size_t x_inc, + const T beta, + cl_mem y_buffer, const size_t y_offset, const size_t y_inc, + cl_command_queue* queue, cl_event* event); + // ================================================================================================= // BLAS level-3 (matrix-matrix) routines diff --git a/include/internal/routines/level2/xsymv.h b/include/internal/routines/level2/xsymv.h new file mode 100644 index 00000000..d6a36489 --- /dev/null +++ b/include/internal/routines/level2/xsymv.h @@ -0,0 +1,58 @@ + +// ================================================================================================= +// 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 Xsymv routine. It is based on the generalized mat-vec multiplication +// routine (Xgemv). The Xsymv class inherits from the templated class Xgemv, allowing it to call the +// "DoGemm" function directly. The "DoSymv" function first preprocesses the symmetric matrix by +// transforming it into a general matrix, and then calls the regular GEMV code. +// +// ================================================================================================= + +#ifndef CLBLAST_ROUTINES_XSYMV_H_ +#define CLBLAST_ROUTINES_XSYMV_H_ + +#include "internal/routines/level2/xgemv.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template +class Xsymv: public Xgemv { + public: + + // Members and methods from the base class + using Routine::db_; + using Routine::context_; + using Routine::GetProgramFromCache; + using Routine::TestMatrixA; + using Routine::RunKernel; + using Routine::ErrorIn; + + // Uses the regular Xgemv routine + using Xgemv::DoGemv; + + // Constructor + Xsymv(Queue &queue, Event &event); + + // Templated-precision implementation of the routine + StatusCode DoSymv(const Layout layout, const Triangle triangle, + const size_t n, + const T alpha, + const Buffer &a_buffer, const size_t a_offset, const size_t a_ld, + const Buffer &x_buffer, const size_t x_offset, const size_t x_inc, + const T beta, + const Buffer &y_buffer, const size_t y_offset, const size_t y_inc); +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_ROUTINES_XSYMV_H_ +#endif diff --git a/src/clblast.cc b/src/clblast.cc index eddb8022..c04462f3 100644 --- a/src/clblast.cc +++ b/src/clblast.cc @@ -22,6 +22,7 @@ // BLAS level-2 includes #include "internal/routines/level2/xgemv.h" +#include "internal/routines/level2/xsymv.h" // BLAS level-3 includes #include "internal/routines/level3/xgemm.h" @@ -36,6 +37,7 @@ namespace clblast { // ================================================================================================= // BLAS level-1 (vector-vector) routines +// ================================================================================================= // AXPY template @@ -75,6 +77,7 @@ template StatusCode Axpy(const size_t, const double2, // ================================================================================================= // BLAS level-2 (matrix-vector) routines +// ================================================================================================= // GEMV template @@ -124,8 +127,47 @@ template StatusCode Gemv(const Layout, const Transpose, cl_mem, const size_t, const size_t, cl_command_queue*, cl_event*); +// ================================================================================================= + +// SYMV +template +StatusCode Symv(const Layout layout, const Triangle triangle, + const size_t n, const T alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, + const cl_mem x_buffer, const size_t x_offset, const size_t x_inc, const T beta, + cl_mem y_buffer, const size_t y_offset, const size_t y_inc, + cl_command_queue* queue, cl_event* event) { + + auto queue_cpp = Queue(*queue); + auto event_cpp = Event(*event); + auto routine = Xsymv(queue_cpp, event_cpp); + + // Compiles the routine's device kernels + auto status = routine.SetUp(); + if (status != StatusCode::kSuccess) { return status; } + + // Runs the routine + return routine.DoSymv(layout, triangle, n, alpha, + Buffer(a_buffer), a_offset, a_ld, + Buffer(x_buffer), x_offset, x_inc, beta, + Buffer(y_buffer), y_offset, y_inc); +} +template StatusCode Symv(const Layout, const Triangle, + const size_t, const float, + const cl_mem, const size_t, const size_t, + 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 Symv(const Layout, const Triangle, + const size_t, const double, + const cl_mem, const size_t, const size_t, + const cl_mem, const size_t, const size_t, const double, + cl_mem, const size_t, const size_t, + cl_command_queue*, cl_event*); + // ================================================================================================= // BLAS level-3 (matrix-matrix) routines +// ================================================================================================= // GEMM template diff --git a/src/routines/level2/xgemv.cc b/src/routines/level2/xgemv.cc index 75219b63..7505e9f7 100644 --- a/src/routines/level2/xgemv.cc +++ b/src/routines/level2/xgemv.cc @@ -30,8 +30,9 @@ template <> const Precision Xgemv::precision_ = Precision::kComplexDoub // Constructor: forwards to base class constructor template Xgemv::Xgemv(Queue &queue, Event &event): - Routine(queue, event, "GEMV", {"Xgemv"}, precision_) { + Routine(queue, event, "GEMV", {"Pad", "Xgemv"}, precision_) { source_string_ = + #include "../../kernels/pad.opencl" // For SymmUpperToSquared and SymmLowerToSquared (for SYMV) #include "../../kernels/xgemv.opencl" ; } diff --git a/src/routines/level2/xsymv.cc b/src/routines/level2/xsymv.cc new file mode 100644 index 00000000..1b0ab6d7 --- /dev/null +++ b/src/routines/level2/xsymv.cc @@ -0,0 +1,100 @@ + +// ================================================================================================= +// 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 Xsymv class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/level2/xsymv.h" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xsymv::Xsymv(Queue &queue, Event &event): + Xgemv(queue, event) { +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xsymv::DoSymv(const Layout layout, const Triangle triangle, + const size_t n, + const T alpha, + const Buffer &a_buffer, const size_t a_offset, const size_t a_ld, + const Buffer &x_buffer, const size_t x_offset, const size_t x_inc, + const T beta, + const Buffer &y_buffer, const size_t y_offset, const size_t y_inc) { + + // Makes sure all dimensions are larger than zero + if (n == 0) { return StatusCode::kInvalidDimension; } + + // Checks for validity of the squared A matrix + auto status = TestMatrixA(n, n, a_buffer, a_offset, a_ld, sizeof(T)); + if (ErrorIn(status)) { return status; } + + // Determines which kernel to run based on the layout (the Xgemv 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_, n*n); + + // Creates a general matrix from the symmetric matrix to be able to run the regular Xgemv + // routine afterwards + try { + auto& program = GetProgramFromCache(); + auto kernel = Kernel(program, kernel_name); + + // Sets the arguments for the symmetric-to-squared kernel + kernel.SetArgument(0, static_cast(n)); + kernel.SetArgument(1, static_cast(a_ld)); + kernel.SetArgument(2, static_cast(a_offset)); + kernel.SetArgument(3, a_buffer()); + kernel.SetArgument(4, static_cast(n)); + kernel.SetArgument(5, static_cast(n)); + 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(n, db_["PAD_WPTX"]), db_["PAD_DIMX"]), + Ceil(CeilDiv(n, db_["PAD_WPTY"]), db_["PAD_DIMY"])}; + auto local = std::vector{db_["PAD_DIMX"], db_["PAD_DIMY"]}; + status = RunKernel(kernel, global, local); + if (ErrorIn(status)) { return status; } + + // Runs the regular Xgemv code + status = DoGemv(layout, Transpose::kNo, n, n, alpha, + temp_symm, 0, n, + x_buffer, x_offset, x_inc, beta, + y_buffer, y_offset, y_inc); + + // Return the status of the Xgemv routine + return status; + } catch (...) { return StatusCode::kInvalidKernel; } + } catch (...) { return StatusCode::kTempBufferAllocFailure; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xsymv; +template class Xsymv; + +// ================================================================================================= +} // namespace clblast diff --git a/test/correctness/routines/level2/xsymv.cc b/test/correctness/routines/level2/xsymv.cc new file mode 100644 index 00000000..a479b999 --- /dev/null +++ b/test/correctness/routines/level2/xsymv.cc @@ -0,0 +1,26 @@ + +// ================================================================================================= +// 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 tests for the Xsymv routine. +// +// ================================================================================================= + +#include "correctness/testblas.h" +#include "routines/level2/xsymv.h" + +// ================================================================================================= + +// Main function (not within the clblast namespace) +int main(int argc, char *argv[]) { + clblast::RunTests, float, float>(argc, argv, false, "SSYMV"); + clblast::RunTests, double, double>(argc, argv, true, "DSYMV"); + return 0; +} + +// ================================================================================================= diff --git a/test/performance/routines/level2/xsymv.cc b/test/performance/routines/level2/xsymv.cc new file mode 100644 index 00000000..30e953a5 --- /dev/null +++ b/test/performance/routines/level2/xsymv.cc @@ -0,0 +1,36 @@ + +// ================================================================================================= +// 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 Xsymv command-line interface performance tester. +// +// ================================================================================================= + +#include "performance/client.h" +#include "routines/level2/xsymv.h" + +// ================================================================================================= + +// 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: + clblast::RunClient, float, float>(argc, argv); break; + case clblast::Precision::kDouble: + clblast::RunClient, double, double>(argc, argv); break; + case clblast::Precision::kComplexSingle: + throw std::runtime_error("Unsupported precision mode"); + case clblast::Precision::kComplexDouble: + throw std::runtime_error("Unsupported precision mode"); + } + return 0; +} + +// ================================================================================================= diff --git a/test/routines/level2/xsymv.h b/test/routines/level2/xsymv.h new file mode 100644 index 00000000..13473a3e --- /dev/null +++ b/test/routines/level2/xsymv.h @@ -0,0 +1,130 @@ + +// ================================================================================================= +// 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 a class with static methods to describe the Xsymv 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_XSYMV_H_ +#define CLBLAST_TEST_ROUTINES_XSYMV_H_ + +#include +#include + +#include "wrapper_clblas.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template +class TestXsymv { + public: + + // The BLAS level: 1, 2, or 3 + static size_t BLASLevel() { return 2; } + + // The list of arguments relevant for this routine + static std::vector GetOptions() { + return {kArgN, + kArgLayout, kArgTriangle, + kArgALeadDim, kArgXInc, kArgYInc, + kArgAOffset, kArgXOffset, kArgYOffset, + kArgAlpha, kArgBeta}; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeX(const Arguments &args) { + return args.n * args.x_inc + args.x_offset; + } + static size_t GetSizeY(const Arguments &args) { + return args.n * args.y_inc + args.y_offset; + } + static size_t GetSizeA(const Arguments &args) { + return args.n * args.a_ld + args.a_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments &args) { + args.a_size = GetSizeA(args); + args.x_size = GetSizeX(args); + args.y_size = GetSizeY(args); + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments &args) { return args.n; } + static size_t DefaultLDB(const Arguments &) { return 1; } // N/A for this routine + static size_t DefaultLDC(const Arguments &) { return 1; } // N/A for this routine + + // Describes which transpose options are relevant for this routine + using Transposes = std::vector; + static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine + static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + + // Describes how to run the CLBlast routine + static StatusCode RunRoutine(const Arguments &args, const Buffers &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = Symv(args.layout, args.triangle, + args.n, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.x_vec(), args.x_offset, args.x_inc, args.beta, + buffers.y_vec(), args.y_offset, args.y_inc, + &queue_plain, &event); + clWaitForEvents(1, &event); + return status; + } + + // Describes how to run the clBLAS routine (for correctness/performance comparison) + static StatusCode RunReference(const Arguments &args, const Buffers &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = clblasXsymv(static_cast(args.layout), + static_cast(args.triangle), + args.n, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.x_vec(), args.x_offset, args.x_inc, args.beta, + buffers.y_vec(), args.y_offset, args.y_inc, + 1, &queue_plain, 0, nullptr, &event); + clWaitForEvents(1, &event); + return static_cast(status); + } + + // Describes how to download the results of the computation (more importantly: which buffer) + static std::vector DownloadResult(const Arguments &args, Buffers &buffers, Queue &queue) { + std::vector result(args.y_size, static_cast(0)); + buffers.y_vec.Read(queue, args.y_size, result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments &args) { + return args.n; + } + static size_t ResultID2(const Arguments &) { return 1; } // N/A for this routine + static size_t GetResultIndex(const Arguments &args, const size_t id1, const size_t) { + return id1*args.y_inc + args.y_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments &args) { + return 2 * args.n * args.n; + } + static size_t GetBytes(const Arguments &args) { + return (args.n*args.n + 2*args.n + args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XSYMV_H_ +#endif diff --git a/test/wrapper_clblas.h b/test/wrapper_clblas.h index 4aaf3705..d69109ac 100644 --- a/test/wrapper_clblas.h +++ b/test/wrapper_clblas.h @@ -132,6 +132,34 @@ clblasStatus clblasXgemv( num_queues, queues, num_wait_events, wait_events, events); } +// Calls {clblasSsymv, clblasDsymv} with the arguments forwarded. +clblasStatus clblasXsymv( + clblasOrder layout, clblasUplo triangle, size_t n, float alpha, + const cl_mem a_mat, size_t a_offset, size_t a_ld, + const cl_mem x_vec, size_t x_offset, size_t x_inc, float beta, + const cl_mem y_vec, size_t y_offset, size_t y_inc, + cl_uint num_queues, cl_command_queue *queues, + cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { + return clblasSsymv(layout, triangle, n, alpha, + a_mat, a_offset, a_ld, + x_vec, x_offset, static_cast(x_inc), beta, + y_vec, y_offset, static_cast(y_inc), + num_queues, queues, num_wait_events, wait_events, events); +} +clblasStatus clblasXsymv( + clblasOrder layout, clblasUplo triangle, size_t n, double alpha, + const cl_mem a_mat, size_t a_offset, size_t a_ld, + const cl_mem x_vec, size_t x_offset, size_t x_inc, double beta, + const cl_mem y_vec, size_t y_offset, size_t y_inc, + cl_uint num_queues, cl_command_queue *queues, + cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { + return clblasDsymv(layout, triangle, n, alpha, + a_mat, a_offset, a_ld, + x_vec, x_offset, static_cast(x_inc), beta, + y_vec, y_offset, static_cast(y_inc), + num_queues, queues, num_wait_events, wait_events, events); +} + // ================================================================================================= // BLAS level-3 (matrix-matrix) routines -- cgit v1.2.3 From 938ca2707fed28cff5ad35d2472c09828ef08ca7 Mon Sep 17 00:00:00 2001 From: CNugteren Date: Fri, 31 Jul 2015 17:35:42 +0200 Subject: Added HEMV routine --- CMakeLists.txt | 2 +- include/clblast.h | 11 +++ include/internal/routines/level2/xhemv.h | 56 +++++++++++++ src/clblast.cc | 39 +++++++++ src/routines/level2/xgemv.cc | 2 +- src/routines/level2/xhemv.cc | 100 +++++++++++++++++++++++ test/correctness/routines/level2/xhemv.cc | 30 +++++++ test/performance/routines/level2/xhemv.cc | 40 +++++++++ test/routines/level2/xhemv.h | 130 ++++++++++++++++++++++++++++++ test/wrapper_clblas.h | 32 ++++++++ 10 files changed, 440 insertions(+), 2 deletions(-) create mode 100644 include/internal/routines/level2/xhemv.h create mode 100644 src/routines/level2/xhemv.cc create mode 100644 test/correctness/routines/level2/xhemv.cc create mode 100644 test/performance/routines/level2/xhemv.cc create mode 100644 test/routines/level2/xhemv.h diff --git a/CMakeLists.txt b/CMakeLists.txt index 45429dfa..8ca225b2 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -105,7 +105,7 @@ include_directories(${clblast_SOURCE_DIR}/include ${OPENCL_INCLUDE_DIRS}) set(KERNELS copy pad transpose padtranspose xaxpy xgemv xgemm) set(SAMPLE_PROGRAMS sgemm) set(LEVEL1_ROUTINES xaxpy) -set(LEVEL2_ROUTINES xgemv xsymv) +set(LEVEL2_ROUTINES xgemv xhemv xsymv) set(LEVEL3_ROUTINES xgemm xsymm xhemm xsyrk xherk xsyr2k xher2k xtrmm) set(ROUTINES ${LEVEL1_ROUTINES} ${LEVEL2_ROUTINES} ${LEVEL3_ROUTINES}) diff --git a/include/clblast.h b/include/clblast.h index 93acc03d..bd0f161c 100644 --- a/include/clblast.h +++ b/include/clblast.h @@ -105,6 +105,17 @@ StatusCode Gemv(const Layout layout, const Transpose a_transpose, cl_mem y_buffer, const size_t y_offset, const size_t y_inc, cl_command_queue* queue, cl_event* event); +// Templated-precision hermitian matrix-vector multiplication: SHEMV/DHEMV +template +StatusCode Hemv(const Layout layout, const Triangle triangle, + const size_t n, + const T alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, + const cl_mem x_buffer, const size_t x_offset, const size_t x_inc, + const T beta, + cl_mem y_buffer, const size_t y_offset, const size_t y_inc, + cl_command_queue* queue, cl_event* event); + // Templated-precision symmetric matrix-vector multiplication: SSYMV/DSYMV template StatusCode Symv(const Layout layout, const Triangle triangle, diff --git a/include/internal/routines/level2/xhemv.h b/include/internal/routines/level2/xhemv.h new file mode 100644 index 00000000..4c80fffe --- /dev/null +++ b/include/internal/routines/level2/xhemv.h @@ -0,0 +1,56 @@ + +// ================================================================================================= +// 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 Xhemv routine. It is based on the generalized matrix multiplication +// routine (Xgemv). The implementation is very similar to the Xsymv routine. +// +// ================================================================================================= + +#ifndef CLBLAST_ROUTINES_XHEMV_H_ +#define CLBLAST_ROUTINES_XHEMV_H_ + +#include "internal/routines/level2/xgemv.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template +class Xhemv: public Xgemv { + public: + + // Members and methods from the base class + using Routine::db_; + using Routine::context_; + using Routine::GetProgramFromCache; + using Routine::TestMatrixA; + using Routine::RunKernel; + using Routine::ErrorIn; + + // Uses the regular Xgemv routine + using Xgemv::DoGemv; + + // Constructor + Xhemv(Queue &queue, Event &event); + + // Templated-precision implementation of the routine + StatusCode DoHemv(const Layout layout, const Triangle triangle, + const size_t n, + const T alpha, + const Buffer &a_buffer, const size_t a_offset, const size_t a_ld, + const Buffer &x_buffer, const size_t x_offset, const size_t x_inc, + const T beta, + const Buffer &y_buffer, const size_t y_offset, const size_t y_inc); +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_ROUTINES_XHEMV_H_ +#endif diff --git a/src/clblast.cc b/src/clblast.cc index c04462f3..12c7b880 100644 --- a/src/clblast.cc +++ b/src/clblast.cc @@ -22,6 +22,7 @@ // BLAS level-2 includes #include "internal/routines/level2/xgemv.h" +#include "internal/routines/level2/xhemv.h" #include "internal/routines/level2/xsymv.h" // BLAS level-3 includes @@ -129,6 +130,44 @@ template StatusCode Gemv(const Layout, const Transpose, // ================================================================================================= +// HEMV +template +StatusCode Hemv(const Layout layout, const Triangle triangle, + const size_t n, const T alpha, + const cl_mem a_buffer, const size_t a_offset, const size_t a_ld, + const cl_mem x_buffer, const size_t x_offset, const size_t x_inc, const T beta, + cl_mem y_buffer, const size_t y_offset, const size_t y_inc, + cl_command_queue* queue, cl_event* event) { + + auto queue_cpp = Queue(*queue); + auto event_cpp = Event(*event); + auto routine = Xhemv(queue_cpp, event_cpp); + + // Compiles the routine's device kernels + auto status = routine.SetUp(); + if (status != StatusCode::kSuccess) { return status; } + + // Runs the routine + return routine.DoHemv(layout, triangle, n, alpha, + Buffer(a_buffer), a_offset, a_ld, + Buffer(x_buffer), x_offset, x_inc, beta, + Buffer(y_buffer), y_offset, y_inc); +} +template StatusCode Hemv(const Layout, const Triangle, + const size_t, const float2, + const cl_mem, const size_t, const size_t, + const cl_mem, const size_t, const size_t, const float2, + cl_mem, const size_t, const size_t, + cl_command_queue*, cl_event*); +template StatusCode Hemv(const Layout, const Triangle, + const size_t, const double2, + const cl_mem, const size_t, const size_t, + const cl_mem, const size_t, const size_t, const double2, + cl_mem, const size_t, const size_t, + cl_command_queue*, cl_event*); + +// ================================================================================================= + // SYMV template StatusCode Symv(const Layout layout, const Triangle triangle, diff --git a/src/routines/level2/xgemv.cc b/src/routines/level2/xgemv.cc index 7505e9f7..99b282b0 100644 --- a/src/routines/level2/xgemv.cc +++ b/src/routines/level2/xgemv.cc @@ -32,7 +32,7 @@ template Xgemv::Xgemv(Queue &queue, Event &event): Routine(queue, event, "GEMV", {"Pad", "Xgemv"}, precision_) { source_string_ = - #include "../../kernels/pad.opencl" // For SymmUpperToSquared and SymmLowerToSquared (for SYMV) + #include "../../kernels/pad.opencl" // For {Herm,Symm}{Upper,Lower}ToSquared (for HEMV/SYMV) #include "../../kernels/xgemv.opencl" ; } diff --git a/src/routines/level2/xhemv.cc b/src/routines/level2/xhemv.cc new file mode 100644 index 00000000..7883b7b9 --- /dev/null +++ b/src/routines/level2/xhemv.cc @@ -0,0 +1,100 @@ + +// ================================================================================================= +// 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 Xhemv class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/level2/xhemv.h" + +#include +#include + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template +Xhemv::Xhemv(Queue &queue, Event &event): + Xgemv(queue, event) { +} + +// ================================================================================================= + +// The main routine +template +StatusCode Xhemv::DoHemv(const Layout layout, const Triangle triangle, + const size_t n, + const T alpha, + const Buffer &a_buffer, const size_t a_offset, const size_t a_ld, + const Buffer &x_buffer, const size_t x_offset, const size_t x_inc, + const T beta, + const Buffer &y_buffer, const size_t y_offset, const size_t y_inc) { + + // Makes sure all dimensions are larger than zero + if (n == 0) { return StatusCode::kInvalidDimension; } + + // Checks for validity of the squared A matrix + auto status = TestMatrixA(n, n, a_buffer, a_offset, a_ld, sizeof(T)); + if (ErrorIn(status)) { return status; } + + // Determines which kernel to run based on the layout (the Xgemv 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_, n*n); + + // Creates a general matrix from the hermitian matrix to be able to run the regular Xgemv + // routine afterwards + try { + auto& program = GetProgramFromCache(); + auto kernel = Kernel(program, kernel_name); + + // Sets the arguments for the hermitian-to-squared kernel + kernel.SetArgument(0, static_cast(n)); + kernel.SetArgument(1, static_cast(a_ld)); + kernel.SetArgument(2, static_cast(a_offset)); + kernel.SetArgument(3, a_buffer()); + kernel.SetArgument(4, static_cast(n)); + kernel.SetArgument(5, static_cast(n)); + 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(n, db_["PAD_WPTX"]), db_["PAD_DIMX"]), + Ceil(CeilDiv(n, db_["PAD_WPTY"]), db_["PAD_DIMY"])}; + auto local = std::vector{db_["PAD_DIMX"], db_["PAD_DIMY"]}; + status = RunKernel(kernel, global, local); + if (ErrorIn(status)) { return status; } + + // Runs the regular Xgemv code + status = DoGemv(layout, Transpose::kNo, n, n, alpha, + temp_herm, 0, n, + x_buffer, x_offset, x_inc, beta, + y_buffer, y_offset, y_inc); + + // Return the status of the Xgemv routine + return status; + } catch (...) { return StatusCode::kInvalidKernel; } + } catch (...) { return StatusCode::kTempBufferAllocFailure; } +} + +// ================================================================================================= + +// Compiles the templated class +template class Xhemv; +template class Xhemv; + +// ================================================================================================= +} // namespace clblast diff --git a/test/correctness/routines/level2/xhemv.cc b/test/correctness/routines/level2/xhemv.cc new file mode 100644 index 00000000..183aebc2 --- /dev/null +++ b/test/correctness/routines/level2/xhemv.cc @@ -0,0 +1,30 @@ + +// ================================================================================================= +// 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 tests for the Xhemv routine. +// +// ================================================================================================= + +#include "correctness/testblas.h" +#include "routines/level2/xhemv.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[]) { + clblast::RunTests, float2, float2>(argc, argv, false, "CHEMV"); + clblast::RunTests, double2, double2>(argc, argv, true, "ZHEMV"); + return 0; +} + +// ================================================================================================= diff --git a/test/performance/routines/level2/xhemv.cc b/test/performance/routines/level2/xhemv.cc new file mode 100644 index 00000000..dd70528e --- /dev/null +++ b/test/performance/routines/level2/xhemv.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 +// +// This file implements the Xhemv command-line interface performance tester. +// +// ================================================================================================= + +#include "performance/client.h" +#include "routines/level2/xhemv.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, float2, float2>(argc, argv); break; + case clblast::Precision::kComplexDouble: + clblast::RunClient, double2, double2>(argc, argv); break; + } + return 0; +} + +// ================================================================================================= diff --git a/test/routines/level2/xhemv.h b/test/routines/level2/xhemv.h new file mode 100644 index 00000000..80e22157 --- /dev/null +++ b/test/routines/level2/xhemv.h @@ -0,0 +1,130 @@ + +// ================================================================================================= +// 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 a class with static methods to describe the Xhemv 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_XHEMV_H_ +#define CLBLAST_TEST_ROUTINES_XHEMV_H_ + +#include +#include + +#include "wrapper_clblas.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template +class TestXhemv { + public: + + // The BLAS level: 1, 2, or 3 + static size_t BLASLevel() { return 2; } + + // The list of arguments relevant for this routine + static std::vector GetOptions() { + return {kArgN, + kArgLayout, kArgTriangle, + kArgALeadDim, kArgXInc, kArgYInc, + kArgAOffset, kArgXOffset, kArgYOffset, + kArgAlpha, kArgBeta}; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeX(const Arguments &args) { + return args.n * args.x_inc + args.x_offset; + } + static size_t GetSizeY(const Arguments &args) { + return args.n * args.y_inc + args.y_offset; + } + static size_t GetSizeA(const Arguments &args) { + return args.n * args.a_ld + args.a_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments &args) { + args.a_size = GetSizeA(args); + args.x_size = GetSizeX(args); + args.y_size = GetSizeY(args); + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments &args) { return args.n; } + static size_t DefaultLDB(const Arguments &) { return 1; } // N/A for this routine + static size_t DefaultLDC(const Arguments &) { return 1; } // N/A for this routine + + // Describes which transpose options are relevant for this routine + using Transposes = std::vector; + static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine + static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + + // Describes how to run the CLBlast routine + static StatusCode RunRoutine(const Arguments &args, const Buffers &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = Hemv(args.layout, args.triangle, + args.n, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.x_vec(), args.x_offset, args.x_inc, args.beta, + buffers.y_vec(), args.y_offset, args.y_inc, + &queue_plain, &event); + clWaitForEvents(1, &event); + return status; + } + + // Describes how to run the clBLAS routine (for correctness/performance comparison) + static StatusCode RunReference(const Arguments &args, const Buffers &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = clblasXhemv(static_cast(args.layout), + static_cast(args.triangle), + args.n, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.x_vec(), args.x_offset, args.x_inc, args.beta, + buffers.y_vec(), args.y_offset, args.y_inc, + 1, &queue_plain, 0, nullptr, &event); + clWaitForEvents(1, &event); + return static_cast(status); + } + + // Describes how to download the results of the computation (more importantly: which buffer) + static std::vector DownloadResult(const Arguments &args, Buffers &buffers, Queue &queue) { + std::vector result(args.y_size, static_cast(0)); + buffers.y_vec.Read(queue, args.y_size, result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments &args) { + return args.n; + } + static size_t ResultID2(const Arguments &) { return 1; } // N/A for this routine + static size_t GetResultIndex(const Arguments &args, const size_t id1, const size_t) { + return id1*args.y_inc + args.y_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments &args) { + return 2 * args.n * args.n; + } + static size_t GetBytes(const Arguments &args) { + return (args.n*args.n + 2*args.n + args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XHEMV_H_ +#endif diff --git a/test/wrapper_clblas.h b/test/wrapper_clblas.h index d69109ac..86810fa2 100644 --- a/test/wrapper_clblas.h +++ b/test/wrapper_clblas.h @@ -132,6 +132,38 @@ clblasStatus clblasXgemv( num_queues, queues, num_wait_events, wait_events, events); } +// Calls {clblasChemv, clblasZhemv} with the arguments forwarded. +clblasStatus clblasXhemv( + clblasOrder layout, clblasUplo triangle, size_t n, float2 alpha, + const cl_mem a_mat, size_t a_offset, size_t a_ld, + const cl_mem x_vec, size_t x_offset, size_t x_inc, float2 beta, + const cl_mem y_vec, size_t y_offset, size_t y_inc, + cl_uint num_queues, cl_command_queue *queues, + cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { + auto cl_alpha = cl_float2{{alpha.real(), alpha.imag()}}; + auto cl_beta = cl_float2{{beta.real(), beta.imag()}}; + return clblasChemv(layout, triangle, n, cl_alpha, + a_mat, a_offset, a_ld, + x_vec, x_offset, static_cast(x_inc), cl_beta, + y_vec, y_offset, static_cast(y_inc), + num_queues, queues, num_wait_events, wait_events, events); +} +clblasStatus clblasXhemv( + clblasOrder layout, clblasUplo triangle, size_t n, double2 alpha, + const cl_mem a_mat, size_t a_offset, size_t a_ld, + const cl_mem x_vec, size_t x_offset, size_t x_inc, double2 beta, + const cl_mem y_vec, size_t y_offset, size_t y_inc, + cl_uint num_queues, cl_command_queue *queues, + cl_uint num_wait_events, const cl_event *wait_events, cl_event *events) { + auto cl_alpha = cl_double2{{alpha.real(), alpha.imag()}}; + auto cl_beta = cl_double2{{beta.real(), beta.imag()}}; + return clblasZhemv(layout, triangle, n, cl_alpha, + a_mat, a_offset, a_ld, + x_vec, x_offset, static_cast(x_inc), cl_beta, + y_vec, y_offset, static_cast(y_inc), + num_queues, queues, num_wait_events, wait_events, events); +} + // Calls {clblasSsymv, clblasDsymv} with the arguments forwarded. clblasStatus clblasXsymv( clblasOrder layout, clblasUplo triangle, size_t n, float alpha, -- cgit v1.2.3 From c52c5f3d35d768ca7f16d1b577ed4c6b320f235e Mon Sep 17 00:00:00 2001 From: CNugteren Date: Fri, 31 Jul 2015 17:41:10 +0200 Subject: Added HEMV and SYMV --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index bf70bb3e..965da95e 100644 --- a/README.md +++ b/README.md @@ -149,10 +149,10 @@ CLBlast is in active development and currently does not support the full set of | ---------|---|---|---|---|---------| | xGEMV | ✔ | ✔ | ✔ | ✔ | | | xGBMV | | | | | | -| xHEMV | - | - | | | | +| xHEMV | - | - | ✔ | ✔ | | | xHBMV | - | - | | | | | xHPMV | - | - | | | | -| xSYMV | | | - | - | | +| xSYMV | ✔ | ✔ | - | - | | | xSBMV | | | - | - | | | xSPMV | | | - | - | | | xTRMV | | | | | | -- cgit v1.2.3 From fc7cd434e15b51ff8d39a0fcc8acba4b861ffd18 Mon Sep 17 00:00:00 2001 From: CNugteren Date: Fri, 31 Jul 2015 17:44:17 +0200 Subject: Added HEMV and SYMV --- CHANGELOG | 3 +++ 1 file changed, 3 insertions(+) diff --git a/CHANGELOG b/CHANGELOG index 67096acc..4a87ecb5 100644 --- a/CHANGELOG +++ b/CHANGELOG @@ -2,6 +2,9 @@ Development version (next release) - Now using the Claduc C++11 interface to OpenCL - Removed clBLAS sources, it should now be installed separately for testing +- Added level-2 routines: + * CHEMV/ZHEMV + * SSYMV/DSYMV Version 0.3.0 - Re-organized test/client infrastructure to avoid code duplication -- cgit v1.2.3 From d1a7cf18ecfee1879d00e3a19ce129ee058dd84f Mon Sep 17 00:00:00 2001 From: CNugteren Date: Mon, 3 Aug 2015 07:37:14 +0200 Subject: Abstracted loading of matrix A for GEMV kernel --- src/kernels/xgemv.opencl | 156 ++++++++++++++++++++++++++++------------------- 1 file changed, 94 insertions(+), 62 deletions(-) diff --git a/src/kernels/xgemv.opencl b/src/kernels/xgemv.opencl index 65061717..5bbf69b9 100644 --- a/src/kernels/xgemv.opencl +++ b/src/kernels/xgemv.opencl @@ -52,6 +52,63 @@ R"( // ================================================================================================= +// Data-widths for the 'fast' kernel +#if VW2 == 1 + typedef real realVF; +#elif VW2 == 2 + typedef real2 realVF; +#elif VW2 == 4 + typedef real4 realVF; +#elif VW2 == 8 + typedef real8 realVF; +#elif VW2 == 16 + typedef real16 realVF; +#endif + +// Data-widths for the 'fast' kernel with rotated matrix +#if VW3 == 1 + typedef real realVFR; +#elif VW3 == 2 + typedef real2 realVFR; +#elif VW3 == 4 + typedef real4 realVFR; +#elif VW3 == 8 + typedef real8 realVFR; +#elif VW3 == 16 + typedef real16 realVFR; +#endif + +// ================================================================================================= +// Defines how to load the input matrix in case of a symmetric matrix +#if defined(ROUTINE_SYMV) + +// ================================================================================================= +// Defines how to load the input matrix in case of a hermetian matrix +#elif defined(ROUTINE_HEMV) + +// ================================================================================================= +// Defines how to load the input matrix in the regular case +#else + +// Loads a scalar input value +inline real LoadMatrixA(const __global real* restrict agm, const int x, const int y, + const int a_ld, const int a_offset) { + return agm[x + a_ld*y + a_offset]; +} +// Loads a vector input value (1/2) +inline realVF LoadMatrixAVF(const __global realVF* restrict agm, const int x, const int y, + const int a_ld) { + return agm[x + a_ld*y]; +} +// Loads a vector input value (2/2): as before, but different data-type +inline realVFR LoadMatrixAVFR(const __global realVFR* restrict agm, const int x, const int y, + const int a_ld) { + return agm[x + a_ld*y]; +} + +#endif +// ================================================================================================= + // Full version of the kernel __attribute__((reqd_work_group_size(WGS1, 1, 1))) __kernel void Xgemv(const int m, const int n, const real alpha, const real beta, @@ -96,7 +153,7 @@ __kernel void Xgemv(const int m, const int n, const real alpha, const real beta, #pragma unroll for (int kl=0; kl 'm' and 'n' are multiples of WGS2 // --> 'a_offset' is 0 @@ -203,42 +247,43 @@ __kernel void XgemvFast(const int m, const int n, const real alpha, const real b #pragma unroll for (int w=0; w 'm' and 'n' are multiples of WGS3 // --> 'a_offset' is 0 @@ -311,7 +343,7 @@ __kernel void XgemvFastRot(const int m, const int n, const real alpha, const rea #pragma unroll for (int w=0; w Date: Tue, 4 Aug 2015 08:15:39 +0200 Subject: Added distinguished names for GEMV inherited HEMV/SYMV --- include/internal/routines/level2/xgemv.h | 2 +- include/internal/routines/level2/xhemv.h | 2 +- include/internal/routines/level2/xsymv.h | 2 +- src/kernels/xgemv.opencl | 10 ---------- src/routines/level2/xgemv.cc | 4 ++-- src/routines/level2/xhemv.cc | 4 ++-- src/routines/level2/xsymv.cc | 4 ++-- 7 files changed, 9 insertions(+), 19 deletions(-) diff --git a/include/internal/routines/level2/xgemv.h b/include/internal/routines/level2/xgemv.h index 5ada9b03..1e120a5e 100644 --- a/include/internal/routines/level2/xgemv.h +++ b/include/internal/routines/level2/xgemv.h @@ -36,7 +36,7 @@ class Xgemv: public Routine { using Routine::ErrorIn; // Constructor - Xgemv(Queue &queue, Event &event); + Xgemv(Queue &queue, Event &event, const std::string &name = "GEMV"); // Templated-precision implementation of the routine StatusCode DoGemv(const Layout layout, const Transpose a_transpose, diff --git a/include/internal/routines/level2/xhemv.h b/include/internal/routines/level2/xhemv.h index 4c80fffe..801b2fc3 100644 --- a/include/internal/routines/level2/xhemv.h +++ b/include/internal/routines/level2/xhemv.h @@ -37,7 +37,7 @@ class Xhemv: public Xgemv { using Xgemv::DoGemv; // Constructor - Xhemv(Queue &queue, Event &event); + Xhemv(Queue &queue, Event &event, const std::string &name = "HEMV"); // Templated-precision implementation of the routine StatusCode DoHemv(const Layout layout, const Triangle triangle, diff --git a/include/internal/routines/level2/xsymv.h b/include/internal/routines/level2/xsymv.h index d6a36489..ab6da6d1 100644 --- a/include/internal/routines/level2/xsymv.h +++ b/include/internal/routines/level2/xsymv.h @@ -39,7 +39,7 @@ class Xsymv: public Xgemv { using Xgemv::DoGemv; // Constructor - Xsymv(Queue &queue, Event &event); + Xsymv(Queue &queue, Event &event, const std::string &name = "SYMV"); // Templated-precision implementation of the routine StatusCode DoSymv(const Layout layout, const Triangle triangle, diff --git a/src/kernels/xgemv.opencl b/src/kernels/xgemv.opencl index 5bbf69b9..1e12dd78 100644 --- a/src/kernels/xgemv.opencl +++ b/src/kernels/xgemv.opencl @@ -78,17 +78,8 @@ R"( typedef real16 realVFR; #endif -// ================================================================================================= -// Defines how to load the input matrix in case of a symmetric matrix -#if defined(ROUTINE_SYMV) - -// ================================================================================================= -// Defines how to load the input matrix in case of a hermetian matrix -#elif defined(ROUTINE_HEMV) - // ================================================================================================= // Defines how to load the input matrix in the regular case -#else // Loads a scalar input value inline real LoadMatrixA(const __global real* restrict agm, const int x, const int y, @@ -106,7 +97,6 @@ inline realVFR LoadMatrixAVFR(const __global realVFR* restrict agm, const int x, return agm[x + a_ld*y]; } -#endif // ================================================================================================= // Full version of the kernel diff --git a/src/routines/level2/xgemv.cc b/src/routines/level2/xgemv.cc index 99b282b0..f95a9957 100644 --- a/src/routines/level2/xgemv.cc +++ b/src/routines/level2/xgemv.cc @@ -29,8 +29,8 @@ template <> const Precision Xgemv::precision_ = Precision::kComplexDoub // Constructor: forwards to base class constructor template -Xgemv::Xgemv(Queue &queue, Event &event): - Routine(queue, event, "GEMV", {"Pad", "Xgemv"}, precision_) { +Xgemv::Xgemv(Queue &queue, Event &event, const std::string &name): + Routine(queue, event, name, {"Pad", "Xgemv"}, precision_) { source_string_ = #include "../../kernels/pad.opencl" // For {Herm,Symm}{Upper,Lower}ToSquared (for HEMV/SYMV) #include "../../kernels/xgemv.opencl" diff --git a/src/routines/level2/xhemv.cc b/src/routines/level2/xhemv.cc index 7883b7b9..2d92e45f 100644 --- a/src/routines/level2/xhemv.cc +++ b/src/routines/level2/xhemv.cc @@ -21,8 +21,8 @@ namespace clblast { // Constructor: forwards to base class constructor template -Xhemv::Xhemv(Queue &queue, Event &event): - Xgemv(queue, event) { +Xhemv::Xhemv(Queue &queue, Event &event, const std::string &name): + Xgemv(queue, event, name) { } // ================================================================================================= diff --git a/src/routines/level2/xsymv.cc b/src/routines/level2/xsymv.cc index 1b0ab6d7..2ccb51f6 100644 --- a/src/routines/level2/xsymv.cc +++ b/src/routines/level2/xsymv.cc @@ -21,8 +21,8 @@ namespace clblast { // Constructor: forwards to base class constructor template -Xsymv::Xsymv(Queue &queue, Event &event): - Xgemv(queue, event) { +Xsymv::Xsymv(Queue &queue, Event &event, const std::string &name): + Xgemv(queue, event, name) { } // ================================================================================================= -- cgit v1.2.3