aboutsummaryrefslogtreecommitdiff
path: root/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp')
-rw-r--r--src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp135
1 files changed, 92 insertions, 43 deletions
diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
index 1d9fe4bc01..423592b79c 100644
--- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
@@ -46,79 +46,114 @@ namespace arm_compute
class Coordinates;
} // namespace arm_compute
-CLGEMMLowpMatrixMultiplyKernel::CLGEMMLowpMatrixMultiplyKernel()
- : _input0(nullptr), _input1(nullptr), _output(nullptr)
+namespace
{
-}
+using ElementsProcessed = Steps;
-void CLGEMMLowpMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, bool is_interleaved_transposed)
+Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, bool is_interleaved_transposed)
{
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8);
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::QASYMM8);
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
-
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::QASYMM8);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
if(!is_interleaved_transposed)
{
- ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1));
+ ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != input1->dimension(1));
}
- _input0 = input0;
- _input1 = input1;
- _output = output;
+ return Status{};
+}
- CLBuildOptions build_opts;
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, bool is_interleaved_transposed,
+ ElementsProcessed &num_elements_processed)
+{
+ unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
+ unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
+
+ Window win{};
+ bool window_changed = false;
+ // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication
if(is_interleaved_transposed)
{
- // Create kernel and set static arguments
- build_opts.add_option("-DCOLS_B=" + support::cpp11::to_string(input1->info()->dimension(0)));
- _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_mm_interleaved_transposed", build_opts.options()));
-
// Configure window
- constexpr unsigned int num_elems_processed_per_iteration_x = 16;
- constexpr unsigned int num_elems_processed_per_iteration_y = 4;
+ num_elems_processed_per_iteration_x = 16;
+ num_elems_processed_per_iteration_y = 4;
constexpr unsigned int num_elems_read_per_iteration_input0 = 4;
constexpr unsigned int num_elems_read_per_iteration_input1 = 16;
- Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
-
- AccessWindowRectangle input0_access(input0->info(), 0, 0, num_elems_read_per_iteration_input0, 1);
- AccessWindowRectangle input1_access(input1->info(), 0, 0, num_elems_read_per_iteration_input1, 1);
- AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
+ win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
- update_window_and_padding(win, input0_access, input1_access, output_access);
+ AccessWindowRectangle input0_access(input0, 0, 0, num_elems_read_per_iteration_input0, 1);
+ AccessWindowRectangle input1_access(input1, 0, 0, num_elems_read_per_iteration_input1, 1);
+ AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
- output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));
+ window_changed = update_window_and_padding(win, input0_access, input1_access, output_access);
- ICLKernel::configure(win);
+ output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
}
else
{
// Special case for 1xN, 2xN, 3xN and 4xN input0 tensor. num_elems_processed_per_iteration_x
- constexpr unsigned int num_elems_processed_per_iteration_x = 16;
- const unsigned int num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->info()->dimension(1)), 4);
-
- build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(input0->info()->dimension(0)));
- build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_X=" + support::cpp11::to_string(num_elems_processed_per_iteration_x));
- build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_Y=" + support::cpp11::to_string(num_elems_processed_per_iteration_y));
-
- _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_mm", build_opts.options()));
+ num_elems_processed_per_iteration_x = 16;
+ num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->dimension(1)), 4);
// Configure window
- Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+ win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
- AccessWindowStatic input0_access(input0->info(), 0, 0, input0->info()->dimension(0), ceil_to_multiple(input0->info()->dimension(1), num_elems_processed_per_iteration_y));
- AccessWindowStatic input1_access(input1->info(), 0, 0, ceil_to_multiple(input1->info()->dimension(0), num_elems_processed_per_iteration_x), input1->info()->dimension(1));
- AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
+ AccessWindowStatic input0_access(input0, 0, 0, input0->dimension(0), ceil_to_multiple(input0->dimension(1), num_elems_processed_per_iteration_y));
+ AccessWindowStatic input1_access(input1, 0, 0, ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x), input1->dimension(1));
+ AccessWindowRectangle output_access(output, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
- update_window_and_padding(win, input0_access, input1_access, output_access);
+ window_changed = update_window_and_padding(win, input0_access, input1_access, output_access);
Coordinates coord;
- coord.set_num_dimensions(output->info()->num_dimensions());
- output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape()));
+ coord.set_num_dimensions(output->num_dimensions());
+ output_access.set_valid_region(win, ValidRegion(coord, output->tensor_shape()));
+ }
+
+ Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+ return std::make_pair(err, win);
+}
+} // namespace
+
+CLGEMMLowpMatrixMultiplyKernel::CLGEMMLowpMatrixMultiplyKernel()
+ : _input0(nullptr), _input1(nullptr), _output(nullptr)
+{
+}
+
+void CLGEMMLowpMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, bool is_interleaved_transposed)
+{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), is_interleaved_transposed));
+
+ _input0 = input0;
+ _input1 = input1;
+ _output = output;
+
+ ElementsProcessed num_elements_processed{};
+
+ // Configure kernel window
+ auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), is_interleaved_transposed, num_elements_processed);
+ ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+ ICLKernel::configure(win_config.second);
- ICLKernel::configure(win);
+ // Create build options
+ CLBuildOptions build_opts;
+ std::string kernel_name(" ");
+ if(is_interleaved_transposed)
+ {
+ build_opts.add_option("-DCOLS_B=" + support::cpp11::to_string(input1->info()->dimension(0)));
+ kernel_name = "gemmlowp_mm_interleaved_transposed";
}
+ else
+ {
+ build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(input0->info()->dimension(0)));
+ build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_X=" + support::cpp11::to_string(num_elements_processed.x()));
+ build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_Y=" + support::cpp11::to_string(num_elements_processed.y()));
+ kernel_name = "gemmlowp_mm";
+ }
+ // Create kernel
+ _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
// Set config_id for enabling LWS tuning
_config_id = "gemmlowp_";
@@ -132,6 +167,20 @@ void CLGEMMLowpMatrixMultiplyKernel::configure(const ICLTensor *input0, const IC
_config_id += (is_interleaved_transposed ? support::cpp11::to_string(input1->info()->dimension(0)) : support::cpp11::to_string(input1->info()->dimension(1)));
}
+Status CLGEMMLowpMatrixMultiplyKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, bool is_interleaved_transposed)
+{
+ ElementsProcessed num_elements_processed{};
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, is_interleaved_transposed));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
+ input1->clone().get(),
+ output->clone().get(),
+ is_interleaved_transposed,
+ num_elements_processed)
+ .first);
+
+ return Status{};
+}
+
void CLGEMMLowpMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue &queue)
{
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-29 09:57:31 +0000