Expressed HER2K as two HERK calls

4 files changed, 54 insertions, 172 deletions

diff --git a/src/routines/level3/xher2k.cpp b/src/routines/level3/xher2k.cpp
index 25ff0333..6547233b 100644
--- a/src/routines/level3/xher2k.cpp
+++ b/src/routines/level3/xher2k.cpp
@@ -22,18 +22,7 @@ namespace clblast {
 // Constructor: forwards to base class constructor
 template <typename T, typename U>
 Xher2k<T,U>::Xher2k(Queue &queue, EventPointer event, const std::string &name):
-    Routine(queue, event, name, {"Copy","Pad","Transpose","Padtranspose","Xgemm"}, PrecisionValue<T>(), {}, {
-    #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"
-    }) {
+    Xherk<T,U>(queue, event, name) {
 }
 
 // =================================================================================================
@@ -48,160 +37,22 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
                           const U beta,
                           const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld) {
 
-  // Makes sure all dimensions are larger than zero
-  if ((n == 0) || (k == 0) ) { throw BLASError(StatusCode::kInvalidDimension); }
-
-  // Determines whether to apply the conjugate transpose to matrix B (argument: no transpose) or
-  // to matrix A (argument: conjugate transpose)
-  auto ab_conjugate = (ab_transpose != Transpose::kNo);
-
-  // Computes whether or not the matrices are transposed in memory. This is based on their layout
-  // (row or column-major) and whether or not they are requested to be pre-transposed.
-  auto ab_rotated = (layout == Layout::kColMajor && ab_conjugate) ||
-                    (layout == Layout::kRowMajor && !ab_conjugate);
-  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 a1_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 &&
-                    ab_rotated == false && ab_conjugate == false;
-  auto a2_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 &&
-                    ab_rotated == false && ab_conjugate == true;
-  auto b1_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 &&
-                    ab_rotated == false && ab_conjugate == false;
-  auto b2_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && b_ld == n_ceiled && b_offset == 0 &&
-                    ab_rotated == false && ab_conjugate == true;
-
-  // Creates the temporary matrices
-  auto a1_temp = (a1_no_temp) ? a_buffer : Buffer<T>(context_, k_ceiled*n_ceiled);
-  auto a2_temp = (a2_no_temp) ? a_buffer : Buffer<T>(context_, k_ceiled*n_ceiled);
-  auto b1_temp = (b1_no_temp) ? b_buffer : Buffer<T>(context_, k_ceiled*n_ceiled);
-  auto b2_temp = (b2_no_temp) ? b_buffer : Buffer<T>(context_, k_ceiled*n_ceiled);
-  auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled);
-
-  // Convert the arguments to complex versions
+  // Runs the first matrix multiplication
+  auto first_herk_event = Event();
   auto complex_beta = T{beta, static_cast<U>(0.0)};
-
-  // Events of all kernels (including pre/post processing kernels)
-  auto eventWaitList = std::vector<Event>();
-  auto emptyEventList = std::vector<Event>();
-
-  // 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 (!a1_no_temp) {
-    auto eventProcessA1 = Event();
-    PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA1.pointer(), emptyEventList,
-                           ab_one, ab_two, a_ld, a_offset, a_buffer,
-                           n_ceiled, k_ceiled, n_ceiled, 0, a1_temp,
-                           ConstantOne<T>(), program_,
-                           true, ab_rotated, ab_conjugate);
-    eventWaitList.push_back(eventProcessA1);
-  }
-  if (!a2_no_temp) {
-    auto eventProcessA2 = Event();
-    PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA2.pointer(), emptyEventList,
-                           ab_one, ab_two, a_ld, a_offset, a_buffer,
-                           n_ceiled, k_ceiled, n_ceiled, 0, a2_temp,
-                           ConstantOne<T>(), program_,
-                           true, ab_rotated, !ab_conjugate);
-    eventWaitList.push_back(eventProcessA2);
-  }
-  if (!b1_no_temp) {
-    auto eventProcessB1 = Event();
-    PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB1.pointer(), emptyEventList,
-                           ab_one, ab_two, b_ld, b_offset, b_buffer,
-                           n_ceiled, k_ceiled, n_ceiled, 0, b1_temp,
-                           ConstantOne<T>(), program_,
-                           true, ab_rotated, ab_conjugate);
-    eventWaitList.push_back(eventProcessB1);
-  }
-  if (!b2_no_temp) {
-    auto eventProcessB2 = Event();
-    PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB2.pointer(), emptyEventList,
-                           ab_one, ab_two, b_ld, b_offset, b_buffer,
-                           n_ceiled, k_ceiled, n_ceiled, 0, b2_temp,
-                           ConstantOne<T>(), program_,
-                           true, ab_rotated, !ab_conjugate);
-    eventWaitList.push_back(eventProcessB2);
-  }
-
-  // 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<T>(), 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<int>(n_ceiled));
-  kernel.SetArgument(1, static_cast<int>(k_ceiled));
-  kernel.SetArgument(2, GetRealArg(alpha));
-  kernel.SetArgument(3, GetRealArg(complex_beta));
-  kernel.SetArgument(4, a1_temp());
-  kernel.SetArgument(5, b2_temp());
-  kernel.SetArgument(6, c_temp());
-
-  // Computes the global and local thread sizes
-  auto global = std::vector<size_t>{
-    (n_ceiled * db_["MDIMC"]) / db_["MWG"],
-    (n_ceiled * db_["NDIMC"]) / db_["NWG"]
-  };
-  auto local = std::vector<size_t>{db_["MDIMC"], db_["NDIMC"]};
-
-  // Launches the kernel
-  auto eventKernel1 = Event();
-  RunKernel(kernel, queue_, device_, global, local, eventKernel1.pointer(), eventWaitList);
-  eventWaitList.push_back(eventKernel1);
+  const auto negated_ab_transpose = (ab_transpose != Transpose::kNo) ? Transpose::kNo : Transpose::kYes;
+  HerkAB(layout, triangle, ab_transpose, negated_ab_transpose, n, k, alpha,
+         a_buffer, a_offset, a_ld, b_buffer, b_offset, b_ld, complex_beta, c_buffer, c_offset, c_ld,
+         first_herk_event.pointer(), false);
+  ;
+  first_herk_event.WaitForCompletion();
 
   // Swaps the arguments for matrices A and B, sets 'beta' to 1, and conjugate alpha
   auto conjugate_alpha = T{alpha.real(), -alpha.imag()};
   auto complex_one = T{static_cast<U>(1.0), static_cast<U>(0.0)};
-  kernel.SetArgument(2, GetRealArg(conjugate_alpha));
-  kernel.SetArgument(3, GetRealArg(complex_one));
-  kernel.SetArgument(4, b1_temp());
-  kernel.SetArgument(5, a2_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<T>(), program_,
-                         false, c_rotated, false, upper, lower, true);
+  HerkAB(layout, triangle, ab_transpose, negated_ab_transpose, n, k, conjugate_alpha,
+         b_buffer, b_offset, b_ld, a_buffer, a_offset, a_ld, complex_one, c_buffer, c_offset, c_ld,
+         event_, true);
 }
 
 // =================================================================================================
diff --git a/src/routines/level3/xher2k.hpp b/src/routines/level3/xher2k.hpp
index acc346e4..49723702 100644
--- a/src/routines/level3/xher2k.hpp
+++ b/src/routines/level3/xher2k.hpp
@@ -17,14 +17,19 @@
 #define CLBLAST_ROUTINES_XHER2K_H_
 
 #include "routine.hpp"
+#include "routines/level3/xherk.hpp"
 
 namespace clblast {
 // =================================================================================================
 
 // See comment at top of file for a description of the class
 template <typename T, typename U>
-class Xher2k: public Routine {
- public:
+class Xher2k: public Xherk<T, U> {
+public:
+
+  // Uses methods and variables the regular Xherk routine
+  using Xherk<T, U>::event_;
+  using Xherk<T, U>::HerkAB;
 
   // Constructor
   Xher2k(Queue &queue, EventPointer event, const std::string &name = "HER2K");
diff --git a/src/routines/level3/xherk.cpp b/src/routines/level3/xherk.cpp
index 09c8aa2d..6912d3a9 100644
--- a/src/routines/level3/xherk.cpp
+++ b/src/routines/level3/xherk.cpp
@@ -48,6 +48,25 @@ void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Tran
                               const U beta,
                               const Buffer<T> &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;
+  const auto complex_alpha = T{alpha, static_cast<U>(0.0)};
+  const auto complex_beta = T{beta, static_cast<U>(0.0)};
+  HerkAB(layout, triangle, a_transpose, b_transpose, n, k, complex_alpha,
+         a_buffer, a_offset, a_ld, b_buffer, b_offset, b_ld, complex_beta, c_buffer, c_offset, c_ld,
+         event_, true);
+}
+
+template <typename T, typename U>
+void Xherk<T,U>::HerkAB(const Layout layout, const Triangle triangle, const Transpose a_transpose, const Transpose b_transpose,
+                        const size_t n, const size_t k,
+                        const T complex_alpha,
+                        const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
+                        const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld,
+                        const T complex_beta,
+                        const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld,
+                        EventPointer final_event, const bool diagonal_to_zero) {
 
   // 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, dummy1, dummy2;
@@ -69,6 +88,7 @@ void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Tran
   //    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(n, n, c_buffer, c_offset, c_ld);
 
   // Calculates the ceiled versions of n and k
@@ -87,17 +107,13 @@ void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Tran
 
   // Determines whether or not temporary matrices are needed
   const auto a_no_temp = Xgemm<T>::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<T>::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<T>::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<T>(context_, a_one_i * a_two_i);
-  auto b_temp = (b_no_temp) ? a_buffer : Buffer<T>(context_, b_one_i * b_two_i);
+  auto b_temp = (b_no_temp) ? b_buffer : Buffer<T>(context_, b_one_i * b_two_i);
   auto c_temp = Buffer<T>(context_, n_ceiled*n_ceiled);
 
-  // Convert the arguments to complex versions
-  auto complex_alpha = T{alpha, static_cast<U>(0.0)};
-  auto complex_beta = T{beta, static_cast<U>(0.0)};
-
   // Events of all kernels (including pre/post processing kernels)
   auto eventWaitList = std::vector<Event>();
   auto emptyEventList = std::vector<Event>();
@@ -117,7 +133,7 @@ void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Tran
   if (!b_no_temp) {
     auto eventProcessB = Event();
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList,
-                           b_one, b_two, a_ld, a_offset, a_buffer,
+                           b_one, b_two, b_ld, b_offset, b_buffer,
                            b_one_i, b_two_i, b_one_i, 0, b_temp,
                            ConstantOne<T>(), program_,
                            true, b_do_transpose, b_conjugate);
@@ -162,11 +178,11 @@ void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Tran
   const auto upper = Xgemm<T>::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<T>(), program_,
-                         false, c_do_transpose, false, upper, lower, true);
+                         false, c_do_transpose, false, upper, lower, diagonal_to_zero);
 }
 
 // =================================================================================================
diff --git a/src/routines/level3/xherk.hpp b/src/routines/level3/xherk.hpp
index 51f29d7e..690e1565 100644
--- a/src/routines/level3/xherk.hpp
+++ b/src/routines/level3/xherk.hpp
@@ -36,6 +36,16 @@ class Xherk: public Routine {
               const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
               const U beta,
               const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld);
+
+  // Helper function to be reused for HER2K
+  void HerkAB(const Layout layout, const Triangle triangle, const Transpose a_transpose, const Transpose b_transpose,
+              const size_t n, const size_t k,
+              const T complex_alpha,
+              const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
+              const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld,
+              const T complex_beta,
+              const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld,
+              EventPointer final_event, const bool diagonal_to_zero);
 };
 
 // =================================================================================================