haraldwolff/CLBlast
1
0
Fork 0
mirror of https://github.com/CNugteren/CLBlast.git synced 2024-07-04 21:36:57 +02:00
Code Issues Packages Projects Releases Wiki Activity

Made FillMatrix and FillVector functions take a configurable local workgroup size

Browse source
This commit is contained in:
Cedric Nugteren 2018-05-27 12:03:32 +02:00
parent 8e28a7699d
commit 53198121ac
5 changed files with 39 additions and 42 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

58
src/routines/common.cpp
View file

@ -77,11 +77,10 @@ void RunKernel(Kernel &kernel, Queue &queue, const Device &device,
// Sets all elements of a matrix to a constant value // Sets all elements of a matrix to a constant value
template <typename T> template <typename T>
void FillMatrix(Queue &queue, const Device &device, 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, EventPointer event, const std::vector<Event> &waitForEvents,
const size_t m, const size_t n, const size_t ld, const size_t offset, const size_t m, const size_t n, const size_t ld, const size_t offset,
const Buffer<T> &dest, const Buffer<T> &dest, const T constant_value, const size_t local_size) {
const T constant_value) {
auto kernel = Kernel(program, "FillMatrix"); auto kernel = Kernel(program, "FillMatrix");
kernel.SetArgument(0, static_cast<int>(m)); kernel.SetArgument(0, static_cast<int>(m));
kernel.SetArgument(1, static_cast<int>(n)); 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(3, static_cast<int>(offset));
kernel.SetArgument(4, dest()); kernel.SetArgument(4, dest());
kernel.SetArgument(5, GetRealArg(constant_value)); kernel.SetArgument(5, GetRealArg(constant_value));
auto local = std::vector<size_t>{16, 1}; auto local = std::vector<size_t>{local_size, 1};
auto global = std::vector<size_t>{Ceil(m, 16), n}; auto global = std::vector<size_t>{Ceil(m, local_size), n};
RunKernel(kernel, queue, device, global, local, event, waitForEvents); RunKernel(kernel, queue, device, global, local, event, waitForEvents);
} }
// Compiles the above function // 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, EventPointer, const std::vector<Event>&, const size_t, const size_t,
const size_t, const size_t, const Buffer<half>&, const half); 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>, const Databases&, template void FillMatrix<float>(Queue&, const Device&, const std::shared_ptr<Program>,
EventPointer, const std::vector<Event>&, const size_t, const size_t, EventPointer, const std::vector<Event>&, const size_t, const size_t,
const size_t, const size_t, const Buffer<float>&, const float); 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>, const Databases&, template void FillMatrix<double>(Queue&, const Device&, const std::shared_ptr<Program>,
EventPointer, const std::vector<Event>&, const size_t, const size_t, EventPointer, const std::vector<Event>&, const size_t, const size_t,
const size_t, const size_t, const Buffer<double>&, const double); 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>, const Databases&, template void FillMatrix<float2>(Queue&, const Device&, const std::shared_ptr<Program>,
EventPointer, const std::vector<Event>&, const size_t, const size_t, EventPointer, const std::vector<Event>&, const size_t, const size_t,
const size_t, const size_t, const Buffer<float2>&, const float2); 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>, const Databases&, template void FillMatrix<double2>(Queue&, const Device&, const std::shared_ptr<Program>,
EventPointer, const std::vector<Event>&, const size_t, const size_t, 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 // Sets all elements of a vector to a constant value
template <typename T> template <typename T>
void FillVector(Queue &queue, const Device &device, 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, EventPointer event, const std::vector<Event> &waitForEvents,
const size_t n, const size_t inc, const size_t offset, const size_t n, const size_t inc, const size_t offset,
const Buffer<T> &dest, const Buffer<T> &dest, const T constant_value, const size_t local_size) {
const T constant_value) {
auto kernel = Kernel(program, "FillVector"); auto kernel = Kernel(program, "FillVector");
kernel.SetArgument(0, static_cast<int>(n)); kernel.SetArgument(0, static_cast<int>(n));
kernel.SetArgument(1, static_cast<int>(inc)); kernel.SetArgument(1, static_cast<int>(inc));
kernel.SetArgument(2, static_cast<int>(offset)); kernel.SetArgument(2, static_cast<int>(offset));
kernel.SetArgument(3, dest()); kernel.SetArgument(3, dest());
kernel.SetArgument(4, GetRealArg(constant_value)); kernel.SetArgument(4, GetRealArg(constant_value));
auto local = std::vector<size_t>{16}; auto local = std::vector<size_t>{local_size};
auto global = std::vector<size_t>{Ceil(n, 16)}; auto global = std::vector<size_t>{Ceil(n, local_size)};
RunKernel(kernel, queue, device, global, local, event, waitForEvents); RunKernel(kernel, queue, device, global, local, event, waitForEvents);
} }
// Compiles the above function // 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, EventPointer, const std::vector<Event>&, const size_t, const size_t,
const size_t, const Buffer<half>&, const half); const size_t, const Buffer<half>&, const half, const size_t);
template void FillVector<float>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&, template void FillVector<float>(Queue&, const Device&, const std::shared_ptr<Program>,
EventPointer, const std::vector<Event>&, const size_t, const size_t, EventPointer, const std::vector<Event>&, const size_t, const size_t,
const size_t, const Buffer<float>&, const float); const size_t, const Buffer<float>&, const float, const size_t);
template void FillVector<double>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&, template void FillVector<double>(Queue&, const Device&, const std::shared_ptr<Program>,
EventPointer, const std::vector<Event>&, const size_t, const size_t, EventPointer, const std::vector<Event>&, const size_t, const size_t,
const size_t, const Buffer<double>&, const double); const size_t, const Buffer<double>&, const double, const size_t);
template void FillVector<float2>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&, template void FillVector<float2>(Queue&, const Device&, const std::shared_ptr<Program>,
EventPointer, const std::vector<Event>&, const size_t, const size_t, EventPointer, const std::vector<Event>&, const size_t, const size_t,
const size_t, const Buffer<float2>&, const float2); const size_t, const Buffer<float2>&, const float2, const size_t);
template void FillVector<double2>(Queue&, const Device&, const std::shared_ptr<Program>, const Databases&, template void FillVector<double2>(Queue&, const Device&, const std::shared_ptr<Program>,
EventPointer, const std::vector<Event>&, const size_t, const size_t, 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 } // namespace clblast

10
src/routines/common.hpp
View file

@ -36,20 +36,18 @@ void RunKernel(Kernel &kernel, Queue &queue, const Device &device,
// Sets all elements of a matrix to a constant value // Sets all elements of a matrix to a constant value
template <typename T> template <typename T>
void FillMatrix(Queue &queue, const Device &device, 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, EventPointer event, const std::vector<Event> &waitForEvents,
const size_t m, const size_t n, const size_t ld, const size_t offset, const size_t m, const size_t n, const size_t ld, const size_t offset,
const Buffer<T> &dest, const Buffer<T> &dest, const T constant_value, const size_t local_size);
const T constant_value);
// Sets all elements of a vector to a constant value // Sets all elements of a vector to a constant value
template <typename T> template <typename T>
void FillVector(Queue &queue, const Device &device, 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, EventPointer event, const std::vector<Event> &waitForEvents,
const size_t n, const size_t inc, const size_t offset, const size_t n, const size_t inc, const size_t offset,
const Buffer<T> &dest, const Buffer<T> &dest, const T constant_value, const size_t local_size);
const T constant_value);
// ================================================================================================= // =================================================================================================

4
src/routines/level2/xtrsv.cpp
View file

@ -102,8 +102,8 @@ void Xtrsv<T>::DoTrsv(const Layout layout, const Triangle triangle,
// Fills the output buffer with zeros // Fills the output buffer with zeros
auto eventWaitList = std::vector<Event>(); auto eventWaitList = std::vector<Event>();
auto fill_vector_event = Event(); auto fill_vector_event = Event();
FillVector(queue_, device_, program_, db_, fill_vector_event.pointer(), eventWaitList, FillVector(queue_, device_, program_, fill_vector_event.pointer(), eventWaitList,
n, x_inc, x_offset, x_buffer, ConstantZero<T>()); n, x_inc, x_offset, x_buffer, ConstantZero<T>(), 16);
fill_vector_event.WaitForCompletion(); fill_vector_event.WaitForCompletion();
// Derives properties based on the arguments // Derives properties based on the arguments

4
src/routines/level3/xtrsm.cpp
View file

@ -105,8 +105,8 @@ void Xtrsm<T>::TrsmColMajor(const Side side, const Triangle triangle,
// Fills the output buffer with zeros // Fills the output buffer with zeros
auto eventWaitList = std::vector<Event>(); auto eventWaitList = std::vector<Event>();
auto fill_matrix_event = Event(); auto fill_matrix_event = Event();
FillMatrix(queue_, device_, program_, db_, fill_matrix_event.pointer(), eventWaitList, FillMatrix(queue_, device_, program_, fill_matrix_event.pointer(), eventWaitList,
x_one, x_two, x_ld, x_offset, x_buffer, ConstantZero<T>()); x_one, x_two, x_ld, x_offset, x_buffer, ConstantZero<T>(), 16);
fill_matrix_event.WaitForCompletion(); fill_matrix_event.WaitForCompletion();
// Inverts the diagonal blocks // Inverts the diagonal blocks

5
src/routines/levelx/xinvert.cpp
View file

@ -75,8 +75,9 @@ void Xinvert<T>::InvertMatrixDiagonalBlocks(const Layout layout, const Triangle
// Fills the output buffer with zeros // Fills the output buffer with zeros
auto event_wait_list = std::vector<Event>(); auto event_wait_list = std::vector<Event>();
auto fill_matrix_event = Event(); auto fill_matrix_event = Event();
FillMatrix(queue_, device_, program_, db_, fill_matrix_event.pointer(), event_wait_list, FillMatrix(queue_, device_, program_, fill_matrix_event.pointer(), event_wait_list,
block_size, num_blocks * block_size, block_size, 0, dest, ConstantZero<T>()); block_size, num_blocks * block_size, block_size, 0, dest, ConstantZero<T>(),
16);
event_wait_list.push_back(fill_matrix_event); event_wait_list.push_back(fill_matrix_event);
// Inverts the diagonal IB by IB inner blocks of the matrix: one block per work-group // Inverts the diagonal IB by IB inner blocks of the matrix: one block per work-group