Added first version of banded matrix-vector multiplication

7 files changed, 243 insertions, 12 deletions

diff --git a/src/clblast.cc b/src/clblast.cc
index a0dd8c70..ad5e354d 100644
--- a/src/clblast.cc
+++ b/src/clblast.cc
@@ -28,6 +28,7 @@
 
 // BLAS level-2 includes
 #include "internal/routines/level2/xgemv.h"
+#include "internal/routines/level2/xgbmv.h"
 #include "internal/routines/level2/xhemv.h"
 #include "internal/routines/level2/xsymv.h"
 
@@ -327,15 +328,26 @@ template StatusCode Gemv<double2>(const Layout, const Transpose,
 
 // General banded matrix-vector multiplication: SGBMV/DGBMV/CGBMV/ZGBMV
 template <typename T>
-StatusCode Gbmv(const Layout, const Transpose,
-                const size_t, const size_t, const size_t, const size_t,
-                const T,
-                const cl_mem, const size_t, const size_t,
-                const cl_mem, const size_t, const size_t,
-                const T,
-                cl_mem, const size_t, const size_t,
-                cl_command_queue*, cl_event*) {
-  return StatusCode::kNotImplemented;
+StatusCode Gbmv(const Layout layout, const Transpose a_transpose,
+                const size_t m, const size_t n, const size_t kl, const size_t ku,
+                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 = Xgbmv<T>(queue_cpp, event_cpp);
+  auto status = routine.SetUp();
+  if (status != StatusCode::kSuccess) { return status; }
+  return routine.DoGbmv(layout, a_transpose,
+                        m, n, kl, ku,
+                        alpha,
+                        Buffer<T>(a_buffer), a_offset, a_ld,
+                        Buffer<T>(x_buffer), x_offset, x_inc,
+                        beta,
+                        Buffer<T>(y_buffer), y_offset, y_inc);
 }
 template StatusCode Gbmv<float>(const Layout, const Transpose,
                                 const size_t, const size_t, const size_t, const size_t,
diff --git a/src/kernels/xgemv.opencl b/src/kernels/level2/xgemv.opencl
index 1e12dd78..1e12dd78 100644
--- a/src/kernels/xgemv.opencl
+++ b/src/kernels/level2/xgemv.opencl
diff --git a/src/kernels/matrix_transforms/gbgemt.opencl b/src/kernels/matrix_transforms/gbgemt.opencl
new file mode 100644
index 00000000..e46e3a59
--- /dev/null
+++ b/src/kernels/matrix_transforms/gbgemt.opencl
@@ -0,0 +1,60 @@
+
+// =================================================================================================
+// 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 contains the general banded (gb) to general (ge) matrix transforms.
+//
+// This kernel uses the matrix-transforms common tuning parameters.
+//
+// =================================================================================================
+
+// Enables loading of this file using the C++ pre-processor's #include (C++11 standard raw string
+// literal). Comment-out this line for syntax-highlighting when developing.
+R"(
+
+// =================================================================================================
+#if defined(ROUTINE_GBMV)
+
+// Kernel to transform a general banded matrix into a general matrix
+__attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
+__kernel void GeneralBandedToGeneral(const int src_one, const int src_two,
+                                     const int src_ld, const int src_offset,
+                                     __global const real* restrict src,
+                                     const int dest_one, const int dest_two,
+                                     const int dest_ld, const int dest_offset,
+                                     __global real* dest,
+                                     const int layout,
+                                     const int kl, const int ku) {
+
+  // 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_two && id_one < dest_one) {
+        real result;
+        SetToZero(result);
+        const int k = ku - id_two + id_one;
+        if ((id_one >= id_two - ku) && (id_one < id_two + kl + 1)) {
+          result = src[id_two*src_ld + k + src_offset];
+        }
+        dest[id_two*dest_ld + id_one + dest_offset] = result;
+      }
+    }
+  }
+}
+
+#endif
+// =================================================================================================
+
+// End of the C++11 raw string literal
+)"
+
+// =================================================================================================
diff --git a/src/kernels/matrix_transforms/transforms.opencl b/src/kernels/matrix_transforms/transforms.opencl
new file mode 100644
index 00000000..01889a13
--- /dev/null
+++ b/src/kernels/matrix_transforms/transforms.opencl
@@ -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 <www.cedricnugteren.nl>
+//
+// This file contains the common functions and parameters specific for matrix-transform kernels.
+//
+// =================================================================================================
+
+// Enables loading of this file using the C++ pre-processor's #include (C++11 standard raw string
+// literal). Comment-out this line for syntax-highlighting when developing.
+R"(
+
+// =================================================================================================
+
+// Parameters set by the tuner or by the database. Here they are given a basic default value in case
+// this kernel file is used outside of the CLBlast library.
+#ifndef PAD_DIMX
+  #define PAD_DIMX 8      // Local workgroup size in the first dimension (x)
+#endif
+#ifndef PAD_DIMY
+  #define PAD_DIMY 8      // Local workgroup size in the second dimension (y)
+#endif
+#ifndef PAD_WPTX
+  #define PAD_WPTX 1      // Work per thread in the first dimension (x)
+#endif
+#ifndef PAD_WPTY
+  #define PAD_WPTY 1      // Work per thread in the second dimension (y)
+#endif
+
+// =================================================================================================
+
+// End of the C++11 raw string literal
+)"
+
+// =================================================================================================
diff --git a/src/routines/level2/xgbmv.cc b/src/routines/level2/xgbmv.cc
new file mode 100644
index 00000000..eac208b3
--- /dev/null
+++ b/src/routines/level2/xgbmv.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 Xgbmv class (see the header for information about the class).
+//
+// =================================================================================================
+
+#include "internal/routines/level2/xgbmv.h"
+
+#include <string>
+#include <vector>
+
+namespace clblast {
+// =================================================================================================
+
+// Constructor: forwards to base class constructor
+template <typename T>
+Xgbmv<T>::Xgbmv(Queue &queue, Event &event, const std::string &name):
+    Xgemv<T>(queue, event, name) {
+}
+
+// =================================================================================================
+
+// The main routine
+template <typename T>
+StatusCode Xgbmv<T>::DoGbmv(const Layout layout, const Transpose a_transpose,
+                            const size_t m, const size_t n, const size_t kl, const size_t ku,
+                            const T alpha,
+                            const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
+                            const Buffer<T> &x_buffer, const size_t x_offset, const size_t x_inc,
+                            const T beta,
+                            const Buffer<T> &y_buffer, const size_t y_offset, const size_t y_inc) {
+
+  // Makes sure all dimensions are larger than zero
+  if (n == 0 || m == 0) { return StatusCode::kInvalidDimension; }
+
+  //
+  auto rotated = (layout == Layout::kRowMajor);
+  auto t_one = (rotated) ? n : m;
+  auto t_two = (rotated) ? m : n;
+  auto a_one = kl+ku+1;
+  auto a_two = (rotated) ? m : n;
+
+  // Checks for validity of the A matrix
+  auto status = StatusCode::kSuccess;
+  if (a_ld < a_one) { return StatusCode::kInvalidLeadDimA; }
+  try {
+    auto required_size = (a_ld*a_two + a_offset)*sizeof(T);
+    auto buffer_size = a_buffer.GetSize();
+    if (buffer_size < required_size) { return StatusCode::kInsufficientMemoryA; }
+  } catch (...) { return StatusCode::kInvalidMatrixA; }
+
+  // Temporary buffer to generalize the input matrix
+  try {
+    auto t_buffer = Buffer<T>(context_, t_one*t_two);
+
+    // Creates a general matrix from the input to be able to run the regular Xgemv routine
+    try {
+      auto& program = GetProgramFromCache();
+      auto kernel = Kernel(program, "GeneralBandedToGeneral");
+
+      // Sets the arguments for the matrix transform kernel
+      kernel.SetArgument(0, static_cast<int>(a_one));
+      kernel.SetArgument(1, static_cast<int>(a_two));
+      kernel.SetArgument(2, static_cast<int>(a_ld));
+      kernel.SetArgument(3, static_cast<int>(a_offset));
+      kernel.SetArgument(4, a_buffer());
+      kernel.SetArgument(5, static_cast<int>(t_one));
+      kernel.SetArgument(6, static_cast<int>(t_two));
+      kernel.SetArgument(7, static_cast<int>(t_one));
+      kernel.SetArgument(8, static_cast<int>(0));
+      kernel.SetArgument(9, t_buffer());
+      kernel.SetArgument(10, static_cast<int>(layout));
+      if (rotated) {
+        kernel.SetArgument(11, static_cast<int>(ku));
+        kernel.SetArgument(12, static_cast<int>(kl));
+      }
+      else {
+        kernel.SetArgument(11, static_cast<int>(kl));
+        kernel.SetArgument(12, static_cast<int>(ku));
+      }
+
+      // Uses the common matrix-transforms thread configuration
+      auto global = std::vector<size_t>{Ceil(CeilDiv(t_one, db_["PAD_WPTX"]), db_["PAD_DIMX"]),
+                                        Ceil(CeilDiv(t_two, 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 Xgemv code
+      status = DoGemv(layout, a_transpose, m, n, alpha,
+                      t_buffer, 0, t_one,
+                      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 Xgbmv<float>;
+template class Xgbmv<double>;
+template class Xgbmv<float2>;
+template class Xgbmv<double2>;
+
+// =================================================================================================
+} // namespace clblast
diff --git a/src/routines/level2/xgemv.cc b/src/routines/level2/xgemv.cc
index f95a9957..e52d2f20 100644
--- a/src/routines/level2/xgemv.cc
+++ b/src/routines/level2/xgemv.cc
@@ -32,8 +32,10 @@ template <typename T>
 Xgemv<T>::Xgemv(Queue &queue, Event &event, const std::string &name):
     Routine<T>(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"
+    #include "../../kernels/pad.opencl" // TODO: replace
+    #include "../../kernels/matrix_transforms/transforms.opencl"
+    #include "../../kernels/matrix_transforms/gbgemt.opencl"
+    #include "../../kernels/level2/xgemv.opencl"
   ;
 }
 
diff --git a/src/tuning/xgemv.cc b/src/tuning/xgemv.cc
index 3d6fe595..6a066518 100644
--- a/src/tuning/xgemv.cc
+++ b/src/tuning/xgemv.cc
@@ -34,7 +34,7 @@ class TuneXgemv {
   static std::string GetSources() {
     return
       #include "../src/kernels/common.opencl"
-      #include "../src/kernels/xgemv.opencl"
+      #include "../src/kernels/level2/xgemv.opencl"
     ;
   }