From dcce23d938d2df568f8aaade772f87743f53e5b1 Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Wed, 18 Apr 2018 20:29:28 +0200 Subject: Expressed SYR2K as two SYRK calls --- src/routines/level3/xsyr2k.cpp | 157 ++++++----------------------------------- src/routines/level3/xsyr2k.hpp | 9 ++- src/routines/level3/xsyrk.cpp | 27 +++++-- src/routines/level3/xsyrk.hpp | 10 +++ 4 files changed, 59 insertions(+), 144 deletions(-) (limited to 'src/routines/level3') diff --git a/src/routines/level3/xsyr2k.cpp b/src/routines/level3/xsyr2k.cpp index d81428af..65e85b81 100644 --- a/src/routines/level3/xsyr2k.cpp +++ b/src/routines/level3/xsyr2k.cpp @@ -12,6 +12,7 @@ // ================================================================================================= #include "routines/level3/xsyr2k.hpp" +#include "routines/level3/xgemm.hpp" #include #include @@ -22,18 +23,7 @@ namespace clblast { // Constructor: forwards to base class constructor template Xsyr2k::Xsyr2k(Queue &queue, EventPointer event, const std::string &name): - Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue(), {}, { - #include "../../kernels/level3/level3.opencl" - #include "../../kernels/level3/copy_fast.opencl" - #include "../../kernels/level3/copy_pad.opencl" - #include "../../kernels/level3/transpose_fast.opencl" - #include "../../kernels/level3/transpose_pad.opencl" - , // separated in multiple parts to prevent C1091 in MSVC 2013 - #include "../../kernels/level3/xgemm_part1.opencl" - #include "../../kernels/level3/xgemm_part2.opencl" - #include "../../kernels/level3/xgemm_part3.opencl" - #include "../../kernels/level3/xgemm_part4.opencl" - }) { + Xsyrk(queue, event, name) { } // ================================================================================================= @@ -41,134 +31,27 @@ Xsyr2k::Xsyr2k(Queue &queue, EventPointer event, const std::string &name): // The main routine template void Xsyr2k::DoSyr2k(const Layout layout, const Triangle triangle, const Transpose ab_transpose, - const size_t n, const size_t k, - const T alpha, - const Buffer &a_buffer, const size_t a_offset, const size_t a_ld, - const Buffer &b_buffer, const size_t b_offset, const size_t b_ld, - const T beta, - const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { - - // Makes sure all dimensions are larger than zero - if ((n == 0) || (k == 0) ) { throw BLASError(StatusCode::kInvalidDimension); } - - // Computes whether or not the matrices are transposed in memory. This is based on their layout - // (row or column-major) and whether or not they are requested to be pre-transposed. - auto ab_rotated = (layout == Layout::kColMajor && ab_transpose != Transpose::kNo) || - (layout == Layout::kRowMajor && ab_transpose == Transpose::kNo); - auto c_rotated = (layout == Layout::kRowMajor); - - // Computes the first and second dimensions of the A and B matrices taking the layout into account - auto ab_one = (ab_rotated) ? k : n; - auto ab_two = (ab_rotated) ? n : k; - - // Tests the matrices (A, B, C) for validity, first from a perspective of the OpenCL buffers and - // their sizes, and then from a perspective of parameter values (e.g. n, k). Tests whether the - // OpenCL buffers are valid and non-zero and whether the OpenCL buffers have sufficient storage - // space. Also tests that the leading dimensions of: - // matrix A cannot be less than N when rotated, or less than K when not-rotated - // matrix B cannot be less than N when rotated, or less than K when not-rotated - // matrix C cannot be less than N - TestMatrixA(ab_one, ab_two, a_buffer, a_offset, a_ld); - TestMatrixB(ab_one, ab_two, b_buffer, b_offset, b_ld); - TestMatrixC(n, n, c_buffer, c_offset, c_ld); - - // Calculates the ceiled versions of n and k - auto n_ceiled = Ceil(Ceil(n, db_["MWG"]), 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"; - - // Determines whether or not temporary matrices are needed - auto a_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 && - ab_rotated == false; - auto b_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 && - ab_rotated == false; - - // Creates the temporary matrices - auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto b_temp = (b_no_temp) ? b_buffer : Buffer(context_, k_ceiled*n_ceiled); - auto c_temp = Buffer(context_, n_ceiled*n_ceiled); - - // Events of all kernels (including pre/post processing kernels) - auto eventWaitList = std::vector(); - auto emptyEventList = std::vector(); - - // Runs the pre-processing kernels. This transposes the matrices A and B, but also pads zeros to - // to fill it up until it reaches a certain multiple of size (kernel parameter dependent). In - // case nothing has to be done, these kernels can be skipped. - if (!a_no_temp) { - auto eventProcessA = Event(); - PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList, - ab_one, ab_two, a_ld, a_offset, a_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, a_temp, - ConstantOne(), program_, - true, ab_rotated, false); - eventWaitList.push_back(eventProcessA); - } - if (!b_no_temp) { - auto eventProcessB = Event(); - PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList, - ab_one, ab_two, b_ld, b_offset, b_buffer, - n_ceiled, k_ceiled, n_ceiled, 0, b_temp, - ConstantOne(), program_, - true, ab_rotated, false); - eventWaitList.push_back(eventProcessB); - } - - // Furthermore, also creates a (possibly padded) copy of matrix C, since it is not allowed to - // modify the other triangle. - auto eventProcessC = Event(); - PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList, - n, n, c_ld, c_offset, c_buffer, - n_ceiled, n_ceiled, n_ceiled, 0, c_temp, - ConstantOne(), program_, - true, c_rotated, false); - eventWaitList.push_back(eventProcessC); - - // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary - auto kernel = Kernel(program_, kernel_name); - - // Sets the kernel arguments - kernel.SetArgument(0, static_cast(n_ceiled)); - kernel.SetArgument(1, static_cast(k_ceiled)); - kernel.SetArgument(2, GetRealArg(alpha)); - kernel.SetArgument(3, GetRealArg(beta)); - kernel.SetArgument(4, a_temp()); - kernel.SetArgument(5, b_temp()); - kernel.SetArgument(6, c_temp()); - - // Computes the global and local thread sizes - auto global = std::vector{ - (n_ceiled * db_["MDIMC"]) / db_["MWG"], - (n_ceiled * db_["NDIMC"]) / db_["NWG"] - }; - auto local = std::vector{db_["MDIMC"], db_["NDIMC"]}; - - // Launches the kernel - auto eventKernel1 = Event(); - RunKernel(kernel, queue_, device_, global, local, eventKernel1.pointer(), eventWaitList); - eventWaitList.push_back(eventKernel1); + const size_t n, const size_t k, + const T alpha, + const Buffer &a_buffer, const size_t a_offset, const size_t a_ld, + const Buffer &b_buffer, const size_t b_offset, const size_t b_ld, + const T beta, + const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { + + // Runs the first matrix multiplication + auto first_syrk_event = Event(); + const auto negated_ab_transpose = (ab_transpose != Transpose::kNo) ? Transpose::kNo : Transpose::kYes; + SyrkAB(layout, triangle, ab_transpose, negated_ab_transpose, n, k, alpha, + a_buffer, a_offset, a_ld, b_buffer, b_offset, b_ld, beta, c_buffer, c_offset, c_ld, + first_syrk_event.pointer()); + ; + first_syrk_event.WaitForCompletion(); // Swaps the arguments for matrices A and B, and sets 'beta' to 1 auto one = ConstantOne(); - kernel.SetArgument(3, GetRealArg(one)); - kernel.SetArgument(4, b_temp()); - kernel.SetArgument(5, a_temp()); - - // Runs the kernel again - auto eventKernel2 = Event(); - RunKernel(kernel, queue_, device_, global, local, eventKernel2.pointer(), eventWaitList); - eventWaitList.push_back(eventKernel2); - - // Runs the post-processing kernel - auto upper = (triangle == Triangle::kUpper); - auto lower = (triangle == Triangle::kLower); - PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList, - n_ceiled, n_ceiled, n_ceiled, 0, c_temp, - n, n, c_ld, c_offset, c_buffer, - ConstantOne(), program_, - false, c_rotated, false, upper, lower, false); + SyrkAB(layout, triangle, ab_transpose, negated_ab_transpose, n, k, alpha, + b_buffer, b_offset, b_ld, a_buffer, a_offset, a_ld, one, c_buffer, c_offset, c_ld, + event_); } // ================================================================================================= diff --git a/src/routines/level3/xsyr2k.hpp b/src/routines/level3/xsyr2k.hpp index a02c6e16..e96bc1ec 100644 --- a/src/routines/level3/xsyr2k.hpp +++ b/src/routines/level3/xsyr2k.hpp @@ -17,14 +17,19 @@ #define CLBLAST_ROUTINES_XSYR2K_H_ #include "routine.hpp" +#include "routines/level3/xsyrk.hpp" namespace clblast { // ================================================================================================= // See comment at top of file for a description of the class template -class Xsyr2k: public Routine { - public: +class Xsyr2k: public Xsyrk { +public: + + // Uses methods and variables the regular Xsyrk routine + using Xsyrk::event_; + using Xsyrk::SyrkAB; // Constructor Xsyr2k(Queue &queue, EventPointer event, const std::string &name = "SYR2K"); diff --git a/src/routines/level3/xsyrk.cpp b/src/routines/level3/xsyrk.cpp index d5771ff8..6bb2a24f 100644 --- a/src/routines/level3/xsyrk.cpp +++ b/src/routines/level3/xsyrk.cpp @@ -48,6 +48,22 @@ void Xsyrk::DoSyrk(const Layout layout, const Triangle triangle, const Transp const T beta, const Buffer &c_buffer, const size_t c_offset, const size_t c_ld) { const auto b_transpose = (a_transpose != Transpose::kNo) ? Transpose::kNo : Transpose::kYes; + const auto b_buffer = a_buffer; + const auto b_offset = a_offset; + const auto b_ld = a_ld; + SyrkAB(layout, triangle, a_transpose, b_transpose, n, k, alpha, + a_buffer, a_offset, a_ld, b_buffer, b_offset, b_ld, beta, c_buffer, c_offset, c_ld, event_); +} + +template +void Xsyrk::SyrkAB(const Layout layout, const Triangle triangle, const Transpose a_transpose, const Transpose b_transpose, + const size_t n, const size_t k, + const T alpha, + const Buffer &a_buffer, const size_t a_offset, const size_t a_ld, + const Buffer &b_buffer, const size_t b_offset, const size_t b_ld, + const T beta, + const Buffer &c_buffer, const size_t c_offset, const size_t c_ld, + EventPointer final_event) { // Computes the transpose/conjugate options and sets the a/b/c sizes based on that bool a_do_transpose, b_do_transpose, c_do_transpose, a_conjugate, b_conjugate; @@ -64,6 +80,7 @@ void Xsyrk::DoSyrk(const Layout layout, const Triangle triangle, const Transp // matrix A cannot be less than N when rotated, or less than K when not-rotated // matrix C cannot be less than N TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld); + TestMatrixB(b_one, b_two, b_buffer, b_offset, b_ld); TestMatrixC(c_one, c_two, c_buffer, c_offset, c_ld); // Calculates the ceiled versions of n and k @@ -82,11 +99,11 @@ void Xsyrk::DoSyrk(const Layout layout, const Triangle triangle, const Transp // Determines whether or not temporary matrices are needed const auto a_no_temp = Xgemm::NoTempBuffer(a_one, a_one_i, a_two, a_two_i, a_ld, a_offset, a_do_transpose, a_conjugate); - const auto b_no_temp = Xgemm::NoTempBuffer(a_one, b_one_i, a_two, b_two_i, a_ld, a_offset, b_do_transpose, b_conjugate); + const auto b_no_temp = Xgemm::NoTempBuffer(b_one, b_one_i, b_two, b_two_i, b_ld, b_offset, b_do_transpose, b_conjugate); // Creates the temporary matrices auto a_temp = (a_no_temp) ? a_buffer : Buffer(context_, a_one_i * a_two_i); - auto b_temp = (b_no_temp) ? a_buffer : Buffer(context_, b_one_i * b_two_i); + auto b_temp = (b_no_temp) ? b_buffer : Buffer(context_, b_one_i * b_two_i); auto c_temp = Buffer(context_, n_ceiled*n_ceiled); // Events of all kernels (including pre/post processing kernels) @@ -108,8 +125,8 @@ void Xsyrk::DoSyrk(const Layout layout, const Triangle triangle, const Transp if (!b_no_temp) { auto eventProcessB = Event(); PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList, - b_one, b_two, a_ld, a_offset, a_buffer, // from A - b_one_i, b_two_i, b_one_i, 0, b_temp, // to a copy, named 'B' + b_one, b_two, b_ld, b_offset, b_buffer, + b_one_i, b_two_i, b_one_i, 0, b_temp, ConstantOne(), program_, true, b_do_transpose, false); eventWaitList.push_back(eventProcessB); @@ -153,7 +170,7 @@ void Xsyrk::DoSyrk(const Layout layout, const Triangle triangle, const Transp const auto upper = Xgemm::c_want_rotated_(db_["GEMMK"]) ? (triangle == Triangle::kLower) : (triangle == Triangle::kUpper); const auto lower = !upper; - PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList, + PadCopyTransposeMatrix(queue_, device_, db_, final_event, eventWaitList, n_ceiled, n_ceiled, n_ceiled, 0, c_temp, n, n, c_ld, c_offset, c_buffer, ConstantOne(), program_, diff --git a/src/routines/level3/xsyrk.hpp b/src/routines/level3/xsyrk.hpp index de42b824..51217ba0 100644 --- a/src/routines/level3/xsyrk.hpp +++ b/src/routines/level3/xsyrk.hpp @@ -38,6 +38,16 @@ class Xsyrk: public Routine { const Buffer &a_buffer, const size_t a_offset, const size_t a_ld, const T beta, const Buffer &c_buffer, const size_t c_offset, const size_t c_ld); + + // Helper function to be reused for SYR2K + void SyrkAB(const Layout layout, const Triangle triangle, const Transpose a_transpose, const Transpose b_transpose, + const size_t n, const size_t k, + const T alpha, + const Buffer &a_buffer, const size_t a_offset, const size_t a_ld, + const Buffer &b_buffer, const size_t b_offset, const size_t b_ld, + const T beta, + const Buffer &c_buffer, const size_t c_offset, const size_t c_ld, + EventPointer final_event); }; // ================================================================================================= -- cgit v1.2.3