Added initial version of GEMV including tester and performance client

5 files changed, 224 insertions, 67 deletions

diff --git a/src/clblast.cc b/src/clblast.cc
index 72de3b24..1d4d0621 100644
--- a/src/clblast.cc
+++ b/src/clblast.cc
@@ -20,6 +20,9 @@
 // BLAS level-1 includes
 #include "internal/routines/xaxpy.h"
 
+// BLAS level-2 includes
+#include "internal/routines/xgemv.h"
+
 // BLAS level-3 includes
 #include "internal/routines/xgemm.h"
 #include "internal/routines/xsymm.h"
@@ -36,18 +39,18 @@ StatusCode Axpy(const size_t n, const T alpha,
                 cl_command_queue* queue, cl_event* event) {
   auto queue_cpp = CommandQueue(*queue);
   auto event_cpp = Event(*event);
-  auto xaxpy = Xaxpy<T>(queue_cpp, event_cpp);
+  auto routine = Xaxpy<T>(queue_cpp, event_cpp);
 
   // Loads the kernel source-code as an include (C++11 raw string literal)
   std::string kernel_source =
   #include "kernels/xaxpy.opencl"
-  auto status = xaxpy.SetUp(kernel_source);
+  auto status = routine.SetUp(kernel_source);
   if (status != StatusCode::kSuccess) { return status; }
 
   // Runs the routine
-  return xaxpy.DoAxpy(n, alpha,
-                      Buffer(x_buffer), x_offset, x_inc,
-                      Buffer(y_buffer), y_offset, y_inc);
+  return routine.DoAxpy(n, alpha,
+                        Buffer(x_buffer), x_offset, x_inc,
+                        Buffer(y_buffer), y_offset, y_inc);
 }
 template StatusCode Axpy<float>(const size_t, const float,
                                 const cl_mem, const size_t, const size_t,
@@ -69,22 +72,70 @@ template StatusCode Axpy<double2>(const size_t, const double2,
 // =================================================================================================
 // BLAS level-2 (matrix-vector) routines
 
+// GEMV
+template <typename T>
+StatusCode Gemv(const Layout layout, const Transpose transpose_a,
+                const size_t m, 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 = CommandQueue(*queue);
+  auto event_cpp = Event(*event);
+  auto routine = Xgemv<T>(queue_cpp, event_cpp);
+
+  // Loads the kernel source-code as an include (C++11 raw string literal)
+  std::string kernel_source =
+  #include "kernels/xgemv.opencl"
+  auto status = routine.SetUp(kernel_source);
+  if (status != StatusCode::kSuccess) { return status; }
+
+  // Runs the routine
+  return routine.DoGemv(layout, transpose_a, m, 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 Gemv<float>(const Layout, const Transpose,
+                                const size_t, 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 Gemv<double>(const Layout, const Transpose,
+                                 const size_t, 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*);
+template StatusCode Gemv<float2>(const Layout, const Transpose,
+                                 const size_t, 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 Gemv<double2>(const Layout, const Transpose,
+                                  const size_t, 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*);
+
 // =================================================================================================
 // BLAS level-3 (matrix-matrix) routines
 
 // GEMM
 template <typename T>
 StatusCode Gemm(const Layout layout, const Transpose transpose_a, const Transpose transpose_b,
-                const size_t m, const size_t n, const size_t k,
-                const T alpha,
+                const size_t m, const size_t n, const size_t k, const T alpha,
                 const cl_mem a_buffer, const size_t a_offset, const size_t a_ld,
-                const cl_mem b_buffer, const size_t b_offset, const size_t b_ld,
-                const T beta,
+                const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const T beta,
                 cl_mem c_buffer, const size_t c_offset, const size_t c_ld,
                 cl_command_queue* queue, cl_event* event) {
   auto queue_cpp = CommandQueue(*queue);
   auto event_cpp = Event(*event);
-  auto xgemm = Xgemm<T>(queue_cpp, event_cpp);
+  auto routine = Xgemm<T>(queue_cpp, event_cpp);
 
   // Loads the kernel source-code as an include (C++11 raw string literal)
   std::string common_source1 =
@@ -97,50 +148,39 @@ StatusCode Gemm(const Layout layout, const Transpose transpose_a, const Transpos
   #include "kernels/padtranspose.opencl"
   std::string kernel_source =
   #include "kernels/xgemm.opencl"
-  auto status = xgemm.SetUp(common_source1 + common_source2 + common_source3 + common_source4 +
-                            kernel_source);
+  auto status = routine.SetUp(common_source1 + common_source2 + common_source3 + common_source4 +
+                              kernel_source);
   if (status != StatusCode::kSuccess) { return status; }
 
   // Runs the routine
-  return xgemm.DoGemm(layout, transpose_a, transpose_b,
-                      m, n, k,
-                      alpha,
-                      Buffer(a_buffer), a_offset, a_ld,
-                      Buffer(b_buffer), b_offset, b_ld,
-                      beta,
-                      Buffer(c_buffer), c_offset, c_ld);
+  return routine.DoGemm(layout, transpose_a, transpose_b, m, n, k, alpha,
+                        Buffer(a_buffer), a_offset, a_ld,
+                        Buffer(b_buffer), b_offset, b_ld, beta,
+                        Buffer(c_buffer), c_offset, c_ld);
 }
 template StatusCode Gemm<float>(const Layout, const Transpose, const Transpose,
-                                const size_t, const size_t, const size_t,
-                                const float,
-                                const cl_mem, const size_t, const size_t,
+                                const size_t, const size_t, const size_t, const float,
                                 const cl_mem, const size_t, const size_t,
-                                const float,
+                                const cl_mem, const size_t, const size_t, const float,
                                 cl_mem, const size_t, const size_t,
                                 cl_command_queue*, cl_event*);
 template StatusCode Gemm<double>(const Layout, const Transpose, const Transpose,
-                                 const size_t, const size_t, const size_t,
-                                 const double,
-                                 const cl_mem, const size_t, const size_t,
+                                 const size_t, const size_t, const size_t, const double,
                                  const cl_mem, const size_t, const size_t,
-                                 const double,
+                                 const cl_mem, const size_t, const size_t, const double,
                                  cl_mem, const size_t, const size_t,
                                  cl_command_queue*, cl_event*);
 /*
 template StatusCode Gemm<float2>(const Layout, const Transpose, const Transpose,
-                                 const size_t, const size_t, const size_t,
-                                 const float2,
+                                 const size_t, const size_t, 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,
+                                 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 Gemm<double2>(const Layout, const Transpose, const Transpose,
-                                  const size_t, const size_t, const size_t,
-                                  const double2,
-                                  const cl_mem, const size_t, const size_t,
+                                  const size_t, const size_t, const size_t, const double2,
                                   const cl_mem, const size_t, const size_t,
-                                  const double2,
+                                  const cl_mem, const size_t, const size_t, const double2,
                                   cl_mem, const size_t, const size_t,
                                   cl_command_queue*, cl_event*);
 */
@@ -150,16 +190,14 @@ template StatusCode Gemm<double2>(const Layout, const Transpose, const Transpose
 // SYMM
 template <typename T>
 StatusCode Symm(const Layout layout, const Side side, const Triangle triangle,
-                const size_t m, const size_t n,
-                const T alpha,
+                const size_t m, 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 b_buffer, const size_t b_offset, const size_t b_ld,
-                const T beta,
+                const cl_mem b_buffer, const size_t b_offset, const size_t b_ld, const T beta,
                 cl_mem c_buffer, const size_t c_offset, const size_t c_ld,
                 cl_command_queue* queue, cl_event* event) {
   auto queue_cpp = CommandQueue(*queue);
   auto event_cpp = Event(*event);
-  auto xsymm = Xsymm<T>(queue_cpp, event_cpp);
+  auto routine = Xsymm<T>(queue_cpp, event_cpp);
 
   // Loads the kernel source-code as an include (C++11 raw string literal)
   std::string common_source1 =
@@ -172,50 +210,39 @@ StatusCode Symm(const Layout layout, const Side side, const Triangle triangle,
   #include "kernels/padtranspose.opencl"
   std::string kernel_source =
   #include "kernels/xgemm.opencl"
-  auto status = xsymm.SetUp(common_source1 + common_source2 + common_source3 + common_source4 +
+  auto status = routine.SetUp(common_source1 + common_source2 + common_source3 + common_source4 +
                             kernel_source);
   if (status != StatusCode::kSuccess) { return status; }
 
   // Runs the routine
-  return xsymm.DoSymm(layout, side, triangle,
-                      m, n,
-                      alpha,
-                      Buffer(a_buffer), a_offset, a_ld,
-                      Buffer(b_buffer), b_offset, b_ld,
-                      beta,
-                      Buffer(c_buffer), c_offset, c_ld);
+  return routine.DoSymm(layout, side, triangle, m, n, alpha,
+                        Buffer(a_buffer), a_offset, a_ld,
+                        Buffer(b_buffer), b_offset, b_ld, beta,
+                        Buffer(c_buffer), c_offset, c_ld);
 }
 template StatusCode Symm<float>(const Layout, const Side, const Triangle,
-                                const size_t, const size_t,
-                                const float,
+                                const size_t, 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,
+                                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 Symm<double>(const Layout, const Side, const Triangle,
-                                 const size_t, const size_t,
-                                 const double,
-                                 const cl_mem, const size_t, const size_t,
+                                 const size_t, const size_t, const double,
                                  const cl_mem, const size_t, const size_t,
-                                 const double,
+                                 const cl_mem, const size_t, const size_t, const double,
                                  cl_mem, const size_t, const size_t,
                                  cl_command_queue*, cl_event*);
 /*
 template StatusCode Symm<float2>(const Layout, const Side, const Triangle,
-                                 const size_t, const size_t,
-                                 const float2,
+                                 const size_t, 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,
+                                 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 Symm<double2>(const Layout, const Side, const Triangle,
-                                  const size_t, const size_t,
-                                  const double2,
-                                  const cl_mem, const size_t, const size_t,
+                                  const size_t, const size_t, const double2,
                                   const cl_mem, const size_t, const size_t,
-                                  const double2,
+                                  const cl_mem, const size_t, const size_t, const double2,
                                   cl_mem, const size_t, const size_t,
                                   cl_command_queue*, cl_event*);
 */
diff --git a/src/database.cc b/src/database.cc
index beaa122b..4d9d844e 100644
--- a/src/database.cc
+++ b/src/database.cc
@@ -13,6 +13,7 @@
 
 #include "internal/database.h"
 #include "internal/database/xaxpy.h"
+#include "internal/database/xgemv.h"
 #include "internal/database/xgemm.h"
 #include "internal/database/copy.h"
 #include "internal/database/pad.h"
@@ -27,6 +28,7 @@ namespace clblast {
 // Initializes the database
 const std::vector<Database::DatabaseEntry> Database::database = {
   XaxpySingle, XaxpyDouble, XaxpyComplexSingle, XaxpyComplexDouble,
+  XgemvSingle, XgemvDouble, XgemvComplexSingle, XgemvComplexDouble,
   XgemmSingle, XgemmDouble, XgemmComplexSingle, XgemmComplexDouble,
   CopySingle, CopyDouble, CopyComplexSingle, CopyComplexDouble,
   PadSingle, PadDouble, PadComplexSingle, PadComplexDouble,
diff --git a/src/kernels/xgemv.opencl b/src/kernels/xgemv.opencl
index c90bc26e..1d2ab435 100644
--- a/src/kernels/xgemv.opencl
+++ b/src/kernels/xgemv.opencl
@@ -34,7 +34,8 @@ R"(
 // The gemv kernel
 __attribute__((reqd_work_group_size(WGS, 1, 1)))
 __kernel void Xgemv(const int m, const int n, const real alpha, const real beta,
-                    const __global real* restrict agm,
+                    const int a_transposed,
+                    const __global real* restrict agm, const int a_offset, const int a_ld,
                     const __global real* restrict xgm, const int x_offset, const int x_inc,
                     __global real* ygm, const int y_offset, const int y_inc) {
 
@@ -45,8 +46,15 @@ __kernel void Xgemv(const int m, const int n, const real alpha, const real beta,
     // Loop over the elements of the matrix A
     real acc;
     SetToZero(acc);
-    for (int k=0; k<n; ++k) {
-      MultiplyAdd(acc, agm[id + m*k], xgm[k*x_inc + x_offset]);
+    if (a_transposed == 0) {
+      for (int k=0; k<n; ++k) {
+        MultiplyAdd(acc, agm[id + a_ld*k + a_offset], xgm[k*x_inc + x_offset]);
+      }
+    }
+    else {
+      for (int k=0; k<n; ++k) {
+        MultiplyAdd(acc, agm[k + a_ld*id + a_offset], xgm[k*x_inc + x_offset]);
+      }
     }
     AXPBY(ygm[id*y_inc + y_offset], alpha, acc, beta, ygm[id*y_inc + y_offset]);
   }
diff --git a/src/routines/xgemv.cc b/src/routines/xgemv.cc
new file mode 100644
index 00000000..67f9536e
--- /dev/null
+++ b/src/routines/xgemv.cc
@@ -0,0 +1,117 @@
+
+// =================================================================================================
+// 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 Xgemv class (see the header for information about the class).
+//
+// =================================================================================================
+
+#include "internal/routines/xgemv.h"
+
+#include <string>
+#include <vector>
+
+namespace clblast {
+// =================================================================================================
+
+// Specific implementations to get the memory-type based on a template argument
+template <> const Precision Xgemv<float>::precision_ = Precision::kSingle;
+template <> const Precision Xgemv<double>::precision_ = Precision::kDouble;
+template <> const Precision Xgemv<float2>::precision_ = Precision::kComplexSingle;
+template <> const Precision Xgemv<double2>::precision_ = Precision::kComplexDouble;
+
+// =================================================================================================
+
+// Constructor: forwards to base class constructor
+template <typename T>
+Xgemv<T>::Xgemv(CommandQueue &queue, Event &event):
+    Routine(queue, event, {"Xgemv"}, precision_) {
+}
+
+// =================================================================================================
+
+// The main routine
+template <typename T>
+StatusCode Xgemv<T>::DoGemv(const Layout layout, const Transpose a_transpose,
+                            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 &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 (m == 0 || n == 0) { return StatusCode::kInvalidDimension; }
+
+  // Computes whether or not the matrix has an alternative layout (row or column-major).
+  auto a_altlayout = (layout == Layout::kRowMajor);
+  auto a_one = (a_altlayout) ? n : m;
+  auto a_two = (a_altlayout) ? m : n;
+
+  // Swap m and n if the matrix is transposed
+  auto a_transposed = (a_transpose == Transpose::kYes);
+  auto m_real = (a_transposed) ? n : m;
+  auto n_real = (a_transposed) ? m : n;
+
+  // Determines whether the kernel needs to perform rotated access ('^' is the XOR operator)
+  auto a_rotated = a_transposed ^ a_altlayout;
+
+  // Tests the matrix and the vectors for validity
+  auto status = TestMatrixA(a_one, a_two, a_buffer, a_offset, a_ld, sizeof(T));
+  if (ErrorIn(status)) { return status; }
+  status = TestVectorX(n_real, x_buffer, x_offset, x_inc, sizeof(T));
+  if (ErrorIn(status)) { return status; }
+  status = TestVectorY(m_real, y_buffer, y_offset, y_inc, sizeof(T));
+  if (ErrorIn(status)) { return status; }
+
+  // Retrieves the Xgemv kernel from the compiled binary
+  try {
+    auto program = GetProgramFromCache();
+    auto kernel = Kernel(program, "Xgemv");
+
+    // Sets the kernel arguments
+    kernel.SetArgument(0, static_cast<int>(m_real));
+    kernel.SetArgument(1, static_cast<int>(n_real));
+    kernel.SetArgument(2, alpha);
+    kernel.SetArgument(3, beta);
+    kernel.SetArgument(4, static_cast<int>(a_rotated));
+    kernel.SetArgument(5, a_buffer());
+    kernel.SetArgument(6, static_cast<int>(a_offset));
+    kernel.SetArgument(7, static_cast<int>(a_ld));
+    kernel.SetArgument(8, x_buffer());
+    kernel.SetArgument(9, static_cast<int>(x_offset));
+    kernel.SetArgument(10, static_cast<int>(x_inc));
+    kernel.SetArgument(11, y_buffer());
+    kernel.SetArgument(12, static_cast<int>(y_offset));
+    kernel.SetArgument(13, static_cast<int>(y_inc));
+
+    // Launches the kernel
+    auto m_ceiled = Ceil(m_real, db_["WGS"]);
+    auto global = std::vector<size_t>{CeilDiv(m_ceiled, db_["WPT"])};
+    auto local = std::vector<size_t>{db_["WGS"]};
+    status = RunKernel(kernel, global, local);
+    if (ErrorIn(status)) { return status; }
+
+    // Waits for all kernels to finish
+    queue_.Finish();
+
+    // Succesfully finished the computation
+    return StatusCode::kSuccess;
+  } catch (...) { return StatusCode::kInvalidKernel; }
+}
+
+// =================================================================================================
+
+// Compiles the templated class
+template class Xgemv<float>;
+template class Xgemv<double>;
+template class Xgemv<float2>;
+template class Xgemv<double2>;
+
+// =================================================================================================
+} // namespace clblast
diff --git a/src/tuning/xgemv.cc b/src/tuning/xgemv.cc
index 1ee7c7bf..74bb77a5 100644
--- a/src/tuning/xgemv.cc
+++ b/src/tuning/xgemv.cc
@@ -55,7 +55,10 @@ void XgemvTune(const Arguments<T> &args,
   tuner.AddArgumentScalar(static_cast<int>(args.n));
   tuner.AddArgumentScalar(args.alpha);
   tuner.AddArgumentScalar(args.beta);
+  tuner.AddArgumentScalar(0);
   tuner.AddArgumentInput(a_mat);
+  tuner.AddArgumentScalar(0);
+  tuner.AddArgumentScalar(static_cast<int>(args.n));
   tuner.AddArgumentInput(x_vec);
   tuner.AddArgumentScalar(0);
   tuner.AddArgumentScalar(1);