diff options
author | CNugteren <web@cedricnugteren.nl> | 2015-07-02 07:16:04 +0200 |
---|---|---|
committer | CNugteren <web@cedricnugteren.nl> | 2015-07-02 07:16:04 +0200 |
commit | d9ea0c47c65ff41da2d213cce8b0ef434e817ec2 (patch) | |
tree | 38ca31239e232f53a07a62797b4c06f106d8e8c1 | |
parent | 500416aa38aa2bd03fa24a0d57ac9a2e00cb4c41 (diff) |
Added the TRMM routine, tester, and client
-rw-r--r-- | CMakeLists.txt | 2 | ||||
-rw-r--r-- | include/internal/routines/xtrmm.h | 58 | ||||
-rw-r--r-- | src/clblast.cc | 4 | ||||
-rw-r--r-- | src/kernels/pad.opencl | 97 | ||||
-rw-r--r-- | src/routines/xtrmm.cc | 135 | ||||
-rw-r--r-- | test/correctness/routines/xtrmm.cc | 96 | ||||
-rw-r--r-- | test/performance/graphs/xtrmm.r | 127 | ||||
-rw-r--r-- | test/performance/routines/xtrmm.cc | 36 | ||||
-rw-r--r-- | test/routines/xtrmm.h | 127 |
9 files changed, 666 insertions, 16 deletions
diff --git a/CMakeLists.txt b/CMakeLists.txt index 60b1aaed..60060961 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -98,7 +98,7 @@ set(SAMPLE_PROGRAMS sgemm) set(ROUTINES xaxpy xgemv - xgemm xsymm xsyrk xsyr2k) + xgemm xsymm xsyrk xsyr2k xtrmm) # ================================================================================================== diff --git a/include/internal/routines/xtrmm.h b/include/internal/routines/xtrmm.h new file mode 100644 index 00000000..af9f0266 --- /dev/null +++ b/include/internal/routines/xtrmm.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 <www.cedricnugteren.nl> +// +// This file implements the Xtrmm routine. The implementation is based on first transforming the +// upper/lower unit/non-unit triangular matrix into a regular matrix and then calling the GEMM +// routine. Therefore, this class inherits from the Xgemm class. +// +// ================================================================================================= + +#ifndef CLBLAST_ROUTINES_XTRMM_H_ +#define CLBLAST_ROUTINES_XTRMM_H_ + +#include "internal/routines/xgemm.h" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T> +class Xtrmm: public Xgemm<T> { + public: + + // Uses several variables from the Routine class + using Routine::db_; + using Routine::context_; + + // Uses several helper functions from the Routine class + using Routine::RunKernel; + using Routine::ErrorIn; + using Routine::TestMatrixA; + using Routine::GetProgramFromCache; + + // Uses the regular Xgemm routine + using Xgemm<T>::DoGemm; + + // Constructor + Xtrmm(CommandQueue &queue, Event &event); + + // Templated-precision implementation of the routine + StatusCode 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); +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_ROUTINES_XTRMM_H_ +#endif diff --git a/src/clblast.cc b/src/clblast.cc index e3ce4d39..299d0a18 100644 --- a/src/clblast.cc +++ b/src/clblast.cc @@ -28,6 +28,7 @@ #include "internal/routines/xsymm.h" #include "internal/routines/xsyrk.h" #include "internal/routines/xsyr2k.h" +#include "internal/routines/xtrmm.h" namespace clblast { // ================================================================================================= @@ -372,7 +373,6 @@ StatusCode Trmm(const Layout layout, const Side side, const Triangle triangle, cl_command_queue* queue, cl_event* event) { auto queue_cpp = CommandQueue(*queue); auto event_cpp = Event(*event); - /* auto routine = Xtrmm<T>(queue_cpp, event_cpp); // Loads the kernel source-code as an include (C++11 raw string literal) @@ -394,8 +394,6 @@ StatusCode Trmm(const Layout layout, const Side side, const Triangle triangle, return routine.DoTrmm(layout, side, triangle, a_transpose, diagonal, m, n, alpha, Buffer(a_buffer), a_offset, a_ld, Buffer(b_buffer), b_offset, b_ld); - */ - return StatusCode::kSuccess; } template StatusCode Trmm<float>(const Layout, const Side, const Triangle, const Transpose, const Diagonal, diff --git a/src/kernels/pad.opencl b/src/kernels/pad.opencl index cce0c746..8294fab7 100644 --- a/src/kernels/pad.opencl +++ b/src/kernels/pad.opencl @@ -135,15 +135,15 @@ __kernel void SymmLowerToSquared(const int src_dim, if (id_two < dest_dim && id_one < dest_dim) { // Loads data from the lower-symmetric matrix - real value; - SetToZero(value); + real result; + SetToZero(result); if (id_two < src_dim && id_one < src_dim) { - if (id_two <= id_one) { value = src[id_two*src_ld + id_one + src_offset]; } - else { value = src[id_one*src_ld + id_two + src_offset]; } + if (id_two <= id_one) { result = src[id_two*src_ld + id_one + src_offset]; } + else { result = src[id_one*src_ld + id_two + src_offset]; } } - // Stores the value in the destination matrix - dest[id_two*dest_ld + id_one + dest_offset] = value; + // Stores the result in the destination matrix + dest[id_two*dest_ld + id_one + dest_offset] = result; } } } @@ -168,15 +168,88 @@ __kernel void SymmUpperToSquared(const int src_dim, if (id_two < dest_dim && id_one < dest_dim) { // Loads data from the upper-symmetric matrix - real value; - SetToZero(value); + real result; + SetToZero(result); if (id_two < src_dim && id_one < src_dim) { - if (id_one <= id_two) { value = src[id_two*src_ld + id_one + src_offset]; } - else { value = src[id_one*src_ld + id_two + src_offset]; } + if (id_one <= id_two) { result = src[id_two*src_ld + id_one + src_offset]; } + else { result = src[id_one*src_ld + id_two + src_offset]; } } - // Stores the value in the destination matrix - dest[id_two*dest_ld + id_one + dest_offset] = value; + // Stores the result in the destination matrix + dest[id_two*dest_ld + id_one + dest_offset] = result; + } + } + } +} + +// ================================================================================================= + +// Kernel to populate a squared triangular matrix, given that the triangle which holds the data is +// stored as the lower-triangle of the input matrix. This uses the padding kernel's parameters. +__attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +__kernel void TrmmLowerToSquared(const int src_dim, + const int src_ld, const int src_offset, + __global const real* restrict src, + const int dest_dim, + const int dest_ld, const int dest_offset, + __global real* dest, + const int unit_diagonal) { + + // Loops over the work per thread in both dimensions + #pragma unroll + for (int w_one=0; w_one<PAD_WPTX; ++w_one) { + const int id_one = (get_group_id(0)*PAD_WPTX + w_one) * PAD_DIMX + get_local_id(0); + #pragma unroll + for (int w_two=0; w_two<PAD_WPTY; ++w_two) { + const int id_two = (get_group_id(1)*PAD_WPTY + w_two) * PAD_DIMY + get_local_id(1); + if (id_two < dest_dim && id_one < dest_dim) { + + // Loads data from the lower-triangular matrix + real result; + SetToZero(result); + if (id_two < src_dim && id_one < src_dim) { + if (id_two <= id_one) { result = src[id_two*src_ld + id_one + src_offset]; } + if (id_two == id_one && unit_diagonal) { SetToOne(result); } + // Else: result is zero + } + + // Stores the result in the destination matrix + dest[id_two*dest_ld + id_one + dest_offset] = result; + } + } + } +} + +// Same as above, but now the matrix' data is stored in the upper-triangle +__attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1))) +__kernel void TrmmUpperToSquared(const int src_dim, + const int src_ld, const int src_offset, + __global const real* restrict src, + const int dest_dim, + const int dest_ld, const int dest_offset, + __global real* dest, + const int unit_diagonal) { + + // Loops over the work per thread in both dimensions + #pragma unroll + for (int w_one=0; w_one<PAD_WPTX; ++w_one) { + const int id_one = (get_group_id(0)*PAD_WPTX + w_one) * PAD_DIMX + get_local_id(0); + #pragma unroll + for (int w_two=0; w_two<PAD_WPTY; ++w_two) { + const int id_two = (get_group_id(1)*PAD_WPTY + w_two) * PAD_DIMY + get_local_id(1); + if (id_two < dest_dim && id_one < dest_dim) { + + // Loads data from the upper-triangular matrix + real result; + SetToZero(result); + if (id_two < src_dim && id_one < src_dim) { + if (id_one <= id_two) { result = src[id_two*src_ld + id_one + src_offset]; } + if (id_one == id_two && unit_diagonal) { SetToOne(result); } + // Else: result is zero + } + + // Stores the result in the destination matrix + dest[id_two*dest_ld + id_one + dest_offset] = result; } } } diff --git a/src/routines/xtrmm.cc b/src/routines/xtrmm.cc new file mode 100644 index 00000000..543df844 --- /dev/null +++ b/src/routines/xtrmm.cc @@ -0,0 +1,135 @@ + +// ================================================================================================= +// 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 Xtrmm class (see the header for information about the class). +// +// ================================================================================================= + +#include "internal/routines/xtrmm.h" + +#include <string> +#include <vector> + +namespace clblast { +// ================================================================================================= + +// Constructor: forwards to base class constructor +template <typename T> +Xtrmm<T>::Xtrmm(CommandQueue &queue, Event &event): + Xgemm<T>(queue, event) { +} + +// ================================================================================================= + +// The main routine +template <typename T> +StatusCode Xtrmm<T>::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, sizeof(T)); + 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) ? "TrmmUpperToSquared" : "TrmmLowerToSquared"; + + // 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_, CL_MEM_READ_WRITE, k*k*sizeof(T)); + + // Creates a general matrix from the triangular matrix to be able to run the regular Xgemm + // routine afterwards + try { + auto& program = GetProgramFromCache(); + auto kernel = Kernel(program, kernel_name); + + // Sets the arguments for the triangular-to-squared kernel + kernel.SetArgument(0, static_cast<int>(k)); + kernel.SetArgument(1, static_cast<int>(a_ld)); + kernel.SetArgument(2, static_cast<int>(a_offset)); + kernel.SetArgument(3, a_buffer()); + kernel.SetArgument(4, static_cast<int>(k)); + kernel.SetArgument(5, static_cast<int>(k)); + kernel.SetArgument(6, static_cast<int>(0)); + kernel.SetArgument(7, temp_triangular()); + kernel.SetArgument(8, static_cast<int>(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<size_t>{Ceil(CeilDiv(k, db_["PAD_WPTX"]), db_["PAD_DIMX"]), + Ceil(CeilDiv(k, db_["PAD_WPTY"]), db_["PAD_DIMY"])}; + auto local = std::vector<size_t>{db_["PAD_DIMX"], db_["PAD_DIMY"]}; + status = RunKernel(kernel, global, local); + if (ErrorIn(status)) { return status; } + + // 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<T>(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<T>(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<float>; +template class Xtrmm<double>; +template class Xtrmm<float2>; +template class Xtrmm<double2>; + +// ================================================================================================= +} // namespace clblast diff --git a/test/correctness/routines/xtrmm.cc b/test/correctness/routines/xtrmm.cc new file mode 100644 index 00000000..943fb664 --- /dev/null +++ b/test/correctness/routines/xtrmm.cc @@ -0,0 +1,96 @@ + +// ================================================================================================= +// 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 tests for the Xtrmm routine. +// +// ================================================================================================= + +#include "correctness/testblas.h" +#include "routines/xtrmm.h" + +namespace clblast { +// ================================================================================================= + +// The correctness tester +template <typename T> +void RunTest(int argc, char *argv[], const bool silent, const std::string &name) { + + // Creates a tester + TestBlas<T> tester{argc, argv, silent, name, TestXtrmm<T>::GetOptions(), + TestXtrmm<T>::RunRoutine, TestXtrmm<T>::RunReference, + TestXtrmm<T>::DownloadResult, TestXtrmm<T>::GetResultIndex, + TestXtrmm<T>::ResultID1, TestXtrmm<T>::ResultID2}; + + // This variable holds the arguments relevant for this routine + auto args = Arguments<T>{}; + + // Loops over the test-cases from a data-layout point of view + for (auto &layout: tester.kLayouts) { args.layout = layout; + for (auto &side: tester.kSides) { args.side = side; + for (auto &triangle: tester.kTriangles) { args.triangle = triangle; + for (auto &a_transpose: tester.kTransposes) { args.a_transpose = a_transpose; + for (auto &diagonal: tester.kDiagonals) { args.diagonal = diagonal; + + // Creates the arguments vector for the regular tests + auto regular_test_vector = std::vector<Arguments<T>>{}; + for (auto &m: tester.kMatrixDims) { args.m = m; + for (auto &n: tester.kMatrixDims) { args.n = n; + for (auto &a_ld: tester.kMatrixDims) { args.a_ld = a_ld; + for (auto &a_offset: tester.kOffsets) { args.a_offset = a_offset; + for (auto &b_ld: tester.kMatrixDims) { args.b_ld = b_ld; + for (auto &b_offset: tester.kOffsets) { args.b_offset = b_offset; + for (auto &alpha: tester.kAlphaValues) { args.alpha = alpha; + args.a_size = TestXtrmm<T>::GetSizeA(args); + args.b_size = TestXtrmm<T>::GetSizeB(args); + if (args.a_size<1 || args.b_size<1) { continue; } + regular_test_vector.push_back(args); + } + } + } + } + } + } + } + + // Creates the arguments vector for the invalid-buffer tests + auto invalid_test_vector = std::vector<Arguments<T>>{}; + args.m = args.n = tester.kBufferSize; + args.a_ld = args.b_ld = tester.kBufferSize; + args.a_offset = args.b_offset = 0; + for (auto &a_size: tester.kMatSizes) { args.a_size = a_size; + for (auto &b_size: tester.kMatSizes) { args.b_size = b_size; + invalid_test_vector.push_back(args); + } + } + + // Runs the tests + const auto case_name = ToString(layout)+" "+ToString(side)+" "+ToString(triangle)+" "+ + ToString(a_transpose)+" "+ToString(diagonal); + tester.TestRegular(regular_test_vector, case_name); + tester.TestInvalid(invalid_test_vector, case_name); + } + } + } + } + } +} + +// ================================================================================================= +} // namespace clblast + +// Main function (not within the clblast namespace) +int main(int argc, char *argv[]) { + clblast::RunTest<float>(argc, argv, false, "STRMM"); + clblast::RunTest<double>(argc, argv, true, "DTRMM"); + clblast::RunTest<clblast::float2>(argc, argv, true, "CTRMM"); + clblast::RunTest<clblast::double2>(argc, argv, true, "ZTRMM"); + return 0; +} + +// ================================================================================================= diff --git a/test/performance/graphs/xtrmm.r b/test/performance/graphs/xtrmm.r new file mode 100644 index 00000000..4a3c2440 --- /dev/null +++ b/test/performance/graphs/xtrmm.r @@ -0,0 +1,127 @@ + +# ================================================================================================== +# This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This +# project 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 performance script for the Xtrmm routine +# +# ================================================================================================== + +# Includes the common functions +args <- commandArgs(trailingOnly = FALSE) +thisfile <- (normalizePath(sub("--file=", "", args[grep("--file=", args)]))) +source(file.path(dirname(thisfile), "common.r")) + +# ================================================================================================== + +# Settings +routine_name <- "xtrmm" +parameters <- c("-m","-n","-layout","-side","-triangle","-transA","-diagonal", + "-num_steps","-step","-runs","-precision") +precision <- 32 + +# Sets the names of the test-cases +test_names <- list( + "multiples of 128", + "multiples of 128 (+1)", + "around m=n=512", + "around m=n=2048", + "layouts and side/triangle (m=n=1024)", + "powers of 2" +) + +# Defines the test-cases +test_values <- list( + list(c(128, 128, 0, 0, 0, 0, 0, 16, 128, num_runs, precision)), + list(c(129, 129, 0, 0, 0, 0, 0, 16, 128, num_runs, precision)), + list(c(512, 512, 0, 0, 0, 0, 0, 16, 1, num_runs, precision)), + list(c(2048, 2048, 0, 0, 0, 0, 0, 16, 1, num_runs, precision)), + list( + c(1024, 1024, 0, 0, 0, 0, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 0, 0, 1, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 0, 1, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 0, 1, 1, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 1, 0, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 1, 0, 1, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 1, 1, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 1, 1, 1, 1, 0, num_runs, precision), + + c(1024, 1024, 0, 1, 0, 0, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 1, 0, 0, 1, 1, 0, num_runs, precision), + c(1024, 1024, 0, 1, 0, 1, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 1, 0, 1, 1, 1, 0, num_runs, precision), + c(1024, 1024, 0, 1, 1, 0, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 1, 1, 0, 1, 1, 0, num_runs, precision), + c(1024, 1024, 0, 1, 1, 1, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 1, 1, 1, 1, 1, 0, num_runs, precision), + + c(1024, 1024, 0, 0, 0, 0, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 0, 0, 1, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 0, 1, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 0, 1, 1, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 1, 0, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 1, 0, 1, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 1, 1, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 1, 1, 1, 1, 0, num_runs, precision), + + c(1024, 1024, 1, 1, 0, 0, 0, 1, 0, num_runs, precision), + c(1024, 1024, 1, 1, 0, 0, 1, 1, 0, num_runs, precision), + c(1024, 1024, 1, 1, 0, 1, 0, 1, 0, num_runs, precision), + c(1024, 1024, 1, 1, 0, 1, 1, 1, 0, num_runs, precision), + c(1024, 1024, 1, 1, 1, 0, 0, 1, 0, num_runs, precision), + c(1024, 1024, 1, 1, 1, 0, 1, 1, 0, num_runs, precision), + c(1024, 1024, 1, 1, 1, 1, 0, 1, 0, num_runs, precision), + c(1024, 1024, 1, 1, 1, 1, 1, 1, 0, num_runs, precision) + ), + list( + c(8, 8, 0, 0, 0, 0, 0, 1, 0, num_runs, precision), + c(16, 16, 0, 0, 0, 0, 0, 1, 0, num_runs, precision), + c(32, 32, 0, 0, 0, 0, 0, 1, 0, num_runs, precision), + c(64, 64, 0, 0, 0, 0, 0, 1, 0, num_runs, precision), + c(128, 128, 0, 0, 0, 0, 0, 1, 0, num_runs, precision), + c(256, 256, 0, 0, 0, 0, 0, 1, 0, num_runs, precision), + c(512, 512, 0, 0, 0, 0, 0, 1, 0, num_runs, precision), + c(1024, 1024, 0, 0, 0, 0, 0, 1, 0, num_runs, precision), + c(2048, 2048, 0, 0, 0, 0, 0, 1, 0, num_runs, precision), + c(4096, 4096, 0, 0, 0, 0, 0, 1, 0, num_runs, precision), + c(8192, 8192, 0, 0, 0, 0, 0, 1, 0, num_runs, precision) + ) +) + +# Defines the x-labels corresponding to the test-cases +test_xlabels <- list( + "matrix sizes (m=n)", + "matrix sizes (m=n)", + "matrix sizes (m=n)", + "matrix sizes (m=n)", + "layout (row/col), side (l/r), triangle (up/lo), transA (n/y), diag (u/nu)", + "matrix sizes (m=n)" +) + +# Defines the x-axis of the test-cases +test_xaxis <- list( + c("m", ""), + c("m", ""), + c("m", ""), + c("m", ""), + list(1:32, c("row,l,up,n,u", "row,l,up,n,nu", "row,l,up,y,u", "row,l,up,y,nu", + "row,r,up,n,u", "row,r,up,n,nu", "row,r,up,y,u", "row,r,up,y,nu", + "row,l,lo,n,u", "row,l,lo,n,nu", "row,l,lo,y,u", "row,l,lo,y,nu", + "row,r,lo,n,u", "row,r,lo,n,nu", "row,r,lo,y,u", "row,r,lo,y,nu", + "col,l,up,n,u", "col,l,up,n,nu", "col,l,up,y,u", "col,l,up,y,nu", + "col,r,up,n,u", "col,r,up,n,nu", "col,r,up,y,u", "col,r,up,y,nu", + "col,l,lo,n,u", "col,l,lo,n,nu", "col,l,lo,y,u", "col,l,lo,y,nu", + "col,r,lo,n,u", "col,r,lo,n,nu", "col,r,lo,y,u", "col,r,lo,y,nu")), + c("m", "x") +) + +# ================================================================================================== + +# Start the script +main(routine_name=routine_name, precision=precision, test_names=test_names, test_values=test_values, + test_xlabels=test_xlabels, test_xaxis=test_xaxis, metric_gflops=TRUE) + +# ==================================================================================================
\ No newline at end of file diff --git a/test/performance/routines/xtrmm.cc b/test/performance/routines/xtrmm.cc new file mode 100644 index 00000000..91dcbd07 --- /dev/null +++ b/test/performance/routines/xtrmm.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 <www.cedricnugteren.nl> +// +// This file implements the Xtrmm command-line interface performance tester. +// +// ================================================================================================= + +#include "performance/client.h" +#include "routines/xtrmm.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<clblast::TestXtrmm<float>, float>(argc, argv); break; + case clblast::Precision::kDouble: + clblast::RunClient<clblast::TestXtrmm<double>, double>(argc, argv); break; + case clblast::Precision::kComplexSingle: + clblast::RunClient<clblast::TestXtrmm<clblast::float2>, clblast::float2>(argc, argv); break; + case clblast::Precision::kComplexDouble: + clblast::RunClient<clblast::TestXtrmm<clblast::double2>, clblast::double2>(argc, argv); break; + } + return 0; +} + +// ================================================================================================= diff --git a/test/routines/xtrmm.h b/test/routines/xtrmm.h new file mode 100644 index 00000000..7b7e7af1 --- /dev/null +++ b/test/routines/xtrmm.h @@ -0,0 +1,127 @@ + +// ================================================================================================= +// 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 Xtrmm 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_XTRMM_H_ +#define CLBLAST_TEST_ROUTINES_XTRMM_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> +class TestXtrmm { + public: + + // The list of arguments relevant for this routine + static std::vector<std::string> GetOptions() { + return {kArgM, kArgN, + kArgLayout, kArgSide, kArgTriangle, kArgATransp, kArgDiagonal, + kArgALeadDim, kArgBLeadDim, + kArgAOffset, kArgBOffset, + kArgAlpha}; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeA(const Arguments<T> &args) { + auto k = (args.side == Side::kLeft) ? args.m : args.n; + return k * args.a_ld + args.a_offset; + } + static size_t GetSizeB(const Arguments<T> &args) { + auto b_rotated = (args.layout == Layout::kRowMajor); + auto b_two = (b_rotated) ? args.m : args.n; + return b_two * args.b_ld + args.b_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments<T> &args) { + args.a_size = GetSizeA(args); + args.b_size = GetSizeB(args); + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments<T> &args) { return args.m; } + static size_t DefaultLDB(const Arguments<T> &args) { return args.n; } + static size_t DefaultLDC(const Arguments<T> &) { return 1; } // N/A for this routine + + // Describes how to run the CLBlast routine + static StatusCode RunRoutine(const Arguments<T> &args, const Buffers &buffers, + CommandQueue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = Trmm(args.layout, args.side, args.triangle, args.a_transpose, args.diagonal, + args.m, args.n, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.b_mat(), args.b_offset, args.b_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<T> &args, const Buffers &buffers, + CommandQueue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = clblasXtrmm(static_cast<clblasOrder>(args.layout), + static_cast<clblasSide>(args.side), + static_cast<clblasUplo>(args.triangle), + static_cast<clblasTranspose>(args.a_transpose), + static_cast<clblasDiag>(args.diagonal), + args.m, args.n, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.b_mat(), args.b_offset, args.b_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<T> &args, Buffers &buffers, + CommandQueue &queue) { + std::vector<T> result(args.b_size, static_cast<T>(0)); + buffers.b_mat.ReadBuffer(queue, args.b_size*sizeof(T), result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments<T> &args) { return args.m; } + static size_t ResultID2(const Arguments<T> &args) { return args.n; } + static size_t GetResultIndex(const Arguments<T> &args, const size_t id1, const size_t id2) { + return (args.layout == Layout::kRowMajor) ? + id1*args.b_ld + id2 + args.b_offset: + id2*args.b_ld + id1 + args.b_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments<T> &args) { + auto k = (args.side == Side::kLeft) ? args.m : args.n; + return args.m * args.n * k; + } + static size_t GetBytes(const Arguments<T> &args) { + auto k = (args.side == Side::kLeft) ? args.m : args.n; + return (k*k + 2*args.m*args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XTRMM_H_ +#endif |