Made FillMatrix and FillVector functions take a configurable local workgroup size

5 files changed, 39 insertions, 42 deletions

diff --git a/src/routines/common.cpp b/src/routines/common.cpp
index 5b80e3f2..e81c7873 100644
--- a/src/routines/common.cpp
+++ b/src/routines/common.cpp
@@ -77,11 +77,10 @@ void RunKernel(Kernel &kernel, Queue &queue, const Device &device,
 // Sets all elements of a matrix to a constant value
 template <typename T>
 void FillMatrix(Queue &queue, const Device &device,
-                const std::shared_ptr<Program> program, const Databases &,
+                const std::shared_ptr<Program> program,
                 EventPointer event, const std::vector<Event> &waitForEvents,
                 const size_t m, const size_t n, const size_t ld, const size_t offset,
-                const Buffer<T> &dest,
-                const T constant_value) {
+                const Buffer<T> &dest, const T constant_value, const size_t local_size) {
   auto kernel = Kernel(program, "FillMatrix");
   kernel.SetArgument(0, static_cast<int>(m));
   kernel.SetArgument(1, static_cast<int>(n));
@@ -89,63 +88,62 @@ void FillMatrix(Queue &queue, const Device &device,
   kernel.SetArgument(3, static_cast<int>(offset));
   kernel.SetArgument(4, dest());
   kernel.SetArgument(5, GetRealArg(constant_value));
-  auto local = std::vector<size_t>{16, 1};
-  auto global = std::vector<size_t>{Ceil(m, 16), n};
+  auto local = std::vector<size_t>{local_size, 1};
+  auto global = std::vector<size_t>{Ceil(m, local_size), n};
   RunKernel(kernel, queue, device, global, local, event, waitForEvents);
 }
 
 // Compiles the above function
-template void FillMatrix<half>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&,
+template void FillMatrix<half>(Queue&, const Device&, const std::shared_ptr<Program>,
                                EventPointer, const std::vector<Event>&, const size_t, const size_t,
-                               const size_t, const size_t, const Buffer<half>&, const half);
-template void FillMatrix<float>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&,
+                               const size_t, const size_t, const Buffer<half>&, const half, const size_t);
+template void FillMatrix<float>(Queue&, const Device&, const std::shared_ptr<Program>,
                                 EventPointer, const std::vector<Event>&, const size_t, const size_t,
-                                const size_t, const size_t, const Buffer<float>&, const float);
-template void FillMatrix<double>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&,
+                                const size_t, const size_t, const Buffer<float>&, const float, const size_t);
+template void FillMatrix<double>(Queue&, const Device&, const std::shared_ptr<Program>,
                                  EventPointer, const std::vector<Event>&, const size_t, const size_t,
-                                 const size_t, const size_t, const Buffer<double>&, const double);
-template void FillMatrix<float2>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&,
+                                 const size_t, const size_t, const Buffer<double>&, const double, const size_t);
+template void FillMatrix<float2>(Queue&, const Device&, const std::shared_ptr<Program>,
                                  EventPointer, const std::vector<Event>&, const size_t, const size_t,
-                                 const size_t, const size_t, const Buffer<float2>&, const float2);
-template void FillMatrix<double2>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&,
+                                 const size_t, const size_t, const Buffer<float2>&, const float2, const size_t);
+template void FillMatrix<double2>(Queue&, const Device&, const std::shared_ptr<Program>,
                                   EventPointer, const std::vector<Event>&, const size_t, const size_t,
-                                  const size_t, const size_t, const Buffer<double2>&, const double2);
+                                  const size_t, const size_t, const Buffer<double2>&, const double2, const size_t);
 
 // Sets all elements of a vector to a constant value
 template <typename T>
 void FillVector(Queue &queue, const Device &device,
-                const std::shared_ptr<Program> program, const Databases &,
+                const std::shared_ptr<Program> program,
                 EventPointer event, const std::vector<Event> &waitForEvents,
                 const size_t n, const size_t inc, const size_t offset,
-                const Buffer<T> &dest,
-                const T constant_value) {
+                const Buffer<T> &dest, const T constant_value, const size_t local_size) {
   auto kernel = Kernel(program, "FillVector");
   kernel.SetArgument(0, static_cast<int>(n));
   kernel.SetArgument(1, static_cast<int>(inc));
   kernel.SetArgument(2, static_cast<int>(offset));
   kernel.SetArgument(3, dest());
   kernel.SetArgument(4, GetRealArg(constant_value));
-  auto local = std::vector<size_t>{16};
-  auto global = std::vector<size_t>{Ceil(n, 16)};
+  auto local = std::vector<size_t>{local_size};
+  auto global = std::vector<size_t>{Ceil(n, local_size)};
   RunKernel(kernel, queue, device, global, local, event, waitForEvents);
 }
 
 // Compiles the above function
-template void FillVector<half>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&,
+template void FillVector<half>(Queue&, const Device&, const std::shared_ptr<Program>,
                                EventPointer, const std::vector<Event>&, const size_t, const size_t,
-                               const size_t, const Buffer<half>&, const half);
-template void FillVector<float>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&,
+                               const size_t, const Buffer<half>&, const half, const size_t);
+template void FillVector<float>(Queue&, const Device&, const std::shared_ptr<Program>,
                                 EventPointer, const std::vector<Event>&, const size_t, const size_t,
-                                const size_t, const Buffer<float>&, const float);
-template void FillVector<double>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&,
+                                const size_t, const Buffer<float>&, const float, const size_t);
+template void FillVector<double>(Queue&, const Device&, const std::shared_ptr<Program>,
                                  EventPointer, const std::vector<Event>&, const size_t, const size_t,
-                                 const size_t, const Buffer<double>&, const double);
-template void FillVector<float2>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&,
+                                 const size_t, const Buffer<double>&, const double, const size_t);
+template void FillVector<float2>(Queue&, const Device&, const std::shared_ptr<Program>,
                                  EventPointer, const std::vector<Event>&, const size_t, const size_t,
-                                 const size_t, const Buffer<float2>&, const float2);
-template void FillVector<double2>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&,
+                                 const size_t, const Buffer<float2>&, const float2, const size_t);
+template void FillVector<double2>(Queue&, const Device&, const std::shared_ptr<Program>,
                                   EventPointer, const std::vector<Event>&, const size_t, const size_t,
-                                  const size_t, const Buffer<double2>&, const double2);
+                                  const size_t, const Buffer<double2>&, const double2, const size_t);
 
 // =================================================================================================
 } // namespace clblast
diff --git a/src/routines/common.hpp b/src/routines/common.hpp
index b909243d..c30a2e0e 100644
--- a/src/routines/common.hpp
+++ b/src/routines/common.hpp
@@ -36,20 +36,18 @@ void RunKernel(Kernel &kernel, Queue &queue, const Device &device,
 // Sets all elements of a matrix to a constant value
 template <typename T>
 void FillMatrix(Queue &queue, const Device &device,
-                const std::shared_ptr<Program> program, const Databases &,
+                const std::shared_ptr<Program> program,
                 EventPointer event, const std::vector<Event> &waitForEvents,
                 const size_t m, const size_t n, const size_t ld, const size_t offset,
-                const Buffer<T> &dest,
-                const T constant_value);
+                const Buffer<T> &dest, const T constant_value, const size_t local_size);
 
 // Sets all elements of a vector to a constant value
 template <typename T>
 void FillVector(Queue &queue, const Device &device,
-                const std::shared_ptr<Program> program, const Databases &,
+                const std::shared_ptr<Program> program,
                 EventPointer event, const std::vector<Event> &waitForEvents,
                 const size_t n, const size_t inc, const size_t offset,
-                const Buffer<T> &dest,
-                const T constant_value);
+                const Buffer<T> &dest, const T constant_value, const size_t local_size);
 
 // =================================================================================================
 
diff --git a/src/routines/level2/xtrsv.cpp b/src/routines/level2/xtrsv.cpp
index 36c33a76..c474db4c 100644
--- a/src/routines/level2/xtrsv.cpp
+++ b/src/routines/level2/xtrsv.cpp
@@ -102,8 +102,8 @@ void Xtrsv<T>::DoTrsv(const Layout layout, const Triangle triangle,
   // Fills the output buffer with zeros
   auto eventWaitList = std::vector<Event>();
   auto fill_vector_event = Event();
-  FillVector(queue_, device_, program_, db_, fill_vector_event.pointer(), eventWaitList,
-             n, x_inc, x_offset, x_buffer, ConstantZero<T>());
+  FillVector(queue_, device_, program_, fill_vector_event.pointer(), eventWaitList,
+             n, x_inc, x_offset, x_buffer, ConstantZero<T>(), 16);
   fill_vector_event.WaitForCompletion();
 
   // Derives properties based on the arguments
diff --git a/src/routines/level3/xtrsm.cpp b/src/routines/level3/xtrsm.cpp
index d622e3bf..255e1cca 100644
--- a/src/routines/level3/xtrsm.cpp
+++ b/src/routines/level3/xtrsm.cpp
@@ -105,8 +105,8 @@ void Xtrsm<T>::TrsmColMajor(const Side side, const Triangle triangle,
   // Fills the output buffer with zeros
   auto eventWaitList = std::vector<Event>();
   auto fill_matrix_event = Event();
-  FillMatrix(queue_, device_, program_, db_, fill_matrix_event.pointer(), eventWaitList,
-             x_one, x_two, x_ld, x_offset, x_buffer, ConstantZero<T>());
+  FillMatrix(queue_, device_, program_, fill_matrix_event.pointer(), eventWaitList,
+             x_one, x_two, x_ld, x_offset, x_buffer, ConstantZero<T>(), 16);
   fill_matrix_event.WaitForCompletion();
 
   // Inverts the diagonal blocks
diff --git a/src/routines/levelx/xinvert.cpp b/src/routines/levelx/xinvert.cpp
index a5ef9e10..d851e0b9 100644
--- a/src/routines/levelx/xinvert.cpp
+++ b/src/routines/levelx/xinvert.cpp
@@ -75,8 +75,9 @@ void Xinvert<T>::InvertMatrixDiagonalBlocks(const Layout layout, const Triangle
   // Fills the output buffer with zeros
   auto event_wait_list = std::vector<Event>();
   auto fill_matrix_event = Event();
-  FillMatrix(queue_, device_, program_, db_, fill_matrix_event.pointer(), event_wait_list,
-             block_size, num_blocks * block_size, block_size, 0, dest, ConstantZero<T>());
+  FillMatrix(queue_, device_, program_, fill_matrix_event.pointer(), event_wait_list,
+             block_size, num_blocks * block_size, block_size, 0, dest, ConstantZero<T>(),
+             16);
   event_wait_list.push_back(fill_matrix_event);
 
   // Inverts the diagonal IB by IB inner blocks of the matrix: one block per work-group