diff options
Diffstat (limited to 'src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp')
-rw-r--r-- | src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp | 142 |
1 files changed, 139 insertions, 3 deletions
diff --git a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp index 286b94ebdc..9fa92bde75 100644 --- a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp +++ b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2018 ARM Limited. + * Copyright (c) 2016-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -38,8 +38,8 @@ #include <set> #include <string> -using namespace arm_compute; - +namespace arm_compute +{ namespace { constexpr unsigned int num_elems_processed_per_iteration = 16; @@ -276,3 +276,139 @@ BorderSize CLPixelWiseMultiplicationKernel::border_size() const const unsigned int border = std::min<unsigned int>(num_elems_processed_per_iteration - 1U, replicateSize); return BorderSize(0, border, 0, 0); } + +namespace +{ +constexpr unsigned int num_elems_processed_per_iteration_complex = 1; + +Status validate_arguments_complex(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 2, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 2, DataType::F32); + + const TensorShape &out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape()); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); + + // Validate in case of configured output + if(output->total_size() > 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 2, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), "Wrong shape for output"); + } + + return Status{}; +} + +std::pair<Status, Window> validate_and_configure_window_complex(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output) +{ + const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input1, *input2); + const TensorShape &out_shape = broadcast_pair.first; + const ValidRegion &valid_region = broadcast_pair.second; + + // Auto initialize output if not initialized + const TensorInfo out_info(out_shape, input1->num_channels(), input1->data_type()); + auto_init_if_empty(*output, out_info); + + Window win = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration_complex)); + Window win_input1 = win.broadcast_if_dimension_le_one(*input1); + Window win_input2 = win.broadcast_if_dimension_le_one(*input2); + + AccessWindowHorizontal input1_access(input1, 0, num_elems_processed_per_iteration_complex); + AccessWindowHorizontal input2_access(input2, 0, num_elems_processed_per_iteration_complex); + AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration_complex); + + bool window_changed = update_window_and_padding(win_input1, input1_access) + || update_window_and_padding(win_input2, input2_access) + || update_window_and_padding(win, output_access); + + output_access.set_valid_region(win, valid_region); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} +} // namespace + +CLComplexPixelWiseMultiplicationKernel::CLComplexPixelWiseMultiplicationKernel() + : _input1(nullptr), _input2(nullptr), _output(nullptr) +{ +} + +void CLComplexPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_complex(input1->info(), input2->info(), output->info())); + + // Configure kernel window + auto win_config = validate_and_configure_window_complex(input1->info(), input2->info(), output->info()); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + + _input1 = input1; + _input2 = input2; + _output = output; + + // Create kernel + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("pixelwise_mul_complex")); + + ICLKernel::configure_internal(win_config.second); +} + +Status CLComplexPixelWiseMultiplicationKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_complex(input1, input2, output)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_complex(input1->clone().get(), input2->clone().get(), output->clone().get()).first); + + return Status{}; +} + +void CLComplexPixelWiseMultiplicationKernel::run(const Window &window, cl::CommandQueue &queue) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); + + const TensorShape &in_shape1 = _input1->info()->tensor_shape(); + const TensorShape &in_shape2 = _input2->info()->tensor_shape(); + const TensorShape &out_shape = _output->info()->tensor_shape(); + + bool can_collapse = true; + if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1) + { + can_collapse = (std::min(in_shape1.num_dimensions(), in_shape2.num_dimensions()) > Window::DimZ); + for(size_t d = Window::DimZ; can_collapse && (d < out_shape.num_dimensions()); ++d) + { + can_collapse = (in_shape1[d] == in_shape2[d]); + } + } + + bool has_collapsed = false; + Window collapsed = can_collapse ? window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &has_collapsed) : window; + + const TensorShape &in_shape1_collapsed = has_collapsed ? in_shape1.collapsed_from(Window::DimZ) : in_shape1; + const TensorShape &in_shape2_collapsed = has_collapsed ? in_shape2.collapsed_from(Window::DimZ) : in_shape2; + + Window slice = collapsed.first_slice_window_3D(); + Window slice_input1 = slice.broadcast_if_dimension_le_one(in_shape1_collapsed); + Window slice_input2 = slice.broadcast_if_dimension_le_one(in_shape2_collapsed); + + do + { + unsigned int idx = 0; + add_3D_tensor_argument(idx, _input1, slice_input1); + add_3D_tensor_argument(idx, _input2, slice_input2); + add_3D_tensor_argument(idx, _output, slice); + enqueue(queue, *this, slice); + + collapsed.slide_window_slice_3D(slice_input1); + collapsed.slide_window_slice_3D(slice_input2); + } + while(collapsed.slide_window_slice_3D(slice)); +} + +BorderSize CLComplexPixelWiseMultiplicationKernel::border_size() const +{ + const unsigned int replicateSize = _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0)); + const unsigned int border = std::min<unsigned int>(num_elems_processed_per_iteration_complex - 1U, replicateSize); + return BorderSize(0, border, 0, 0); +} +} // namespace arm_compute
\ No newline at end of file |