From 894066de8cc26d1a3aca62dcaa6b30a2a1116028 Mon Sep 17 00:00:00 2001 From: George Wort Date: Fri, 15 Feb 2019 15:12:52 +0000 Subject: COMPMID-1844: Implement CLCrop Change-Id: I8822c37adc45960705dc3f32a53214795ba3cf39 Signed-off-by: George Wort Reviewed-on: https://review.mlplatform.org/c/789 Reviewed-by: Manuel Bottini Tested-by: Arm Jenkins Reviewed-by: Pablo Marquez --- src/core/CL/kernels/CLMemsetKernel.cpp | 47 ++++++++++++++++++++++------------ 1 file changed, 31 insertions(+), 16 deletions(-) (limited to 'src/core/CL/kernels/CLMemsetKernel.cpp') diff --git a/src/core/CL/kernels/CLMemsetKernel.cpp b/src/core/CL/kernels/CLMemsetKernel.cpp index ab53897543..80caf9406e 100644 --- a/src/core/CL/kernels/CLMemsetKernel.cpp +++ b/src/core/CL/kernels/CLMemsetKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -35,27 +35,38 @@ namespace arm_compute { CLMemsetKernel::CLMemsetKernel() - : ICLKernel(), _tensor(nullptr) + : ICLKernel(), _tensor(nullptr), _full_window() { } void CLMemsetKernel::configure(ICLTensor *tensor, - const PixelValue &constant_value) + const PixelValue &constant_value, + Window *window) { ARM_COMPUTE_ERROR_ON_NULLPTR(tensor); + ARM_COMPUTE_ERROR_THROW_ON(validate(tensor->info(), constant_value, window)); + _tensor = tensor; - const DataType data_type = tensor->info()->data_type(); - const int vec_size_x = 16 / tensor->info()->element_size(); - const int output_width_x = tensor->info()->tensor_shape().x(); - const bool multi_access_x = (output_width_x / vec_size_x > 0); + const DataType data_type = tensor->info()->data_type(); + const int vec_size_x = 16 / tensor->info()->element_size(); // Create and update the window (if needed) - Window win = calculate_max_window(*tensor->info()); + _full_window = calculate_max_window(*tensor->info()); + Window win = _full_window; + if(window != nullptr) + { + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(win, *window); + win = *window; + } + + const int output_width_x = win.num_iterations(0); + const bool multi_access_x = output_width_x >= vec_size_x; + const bool remainder_x = output_width_x % vec_size_x > 0; + if(multi_access_x) { - win.set(Window::DimX, - Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x)); + win.set(Window::DimX, Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x)); } ICLKernel::configure_internal(win); @@ -64,14 +75,18 @@ void CLMemsetKernel::configure(ICLTensor *tensor, build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); build_opts.add_option("-DCONSTANT_VALUE=" + string_from_pixel_value(constant_value, data_type)); build_opts.add_option_if(multi_access_x, "-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); - build_opts.add_option_if(multi_access_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max(output_width_x - vec_size_x, 0))); + build_opts.add_option_if(multi_access_x && remainder_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max(output_width_x - vec_size_x, 0))); _kernel = static_cast(CLKernelLibrary::get().create_kernel("memset", build_opts.options())); } -Status CLMemsetKernel::validate(const ITensorInfo *tensor, const PixelValue &constant_value) +Status CLMemsetKernel::validate(const ITensorInfo *tensor, const PixelValue &constant_value, Window *window) { ARM_COMPUTE_UNUSED(tensor); ARM_COMPUTE_UNUSED(constant_value); + if(window != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON(window->x().step() != 1); + } return Status{}; } @@ -81,15 +96,15 @@ void CLMemsetKernel::run(const Window &window, cl::CommandQueue &queue) ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); // Collapse all the batches on the third - Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimY); - Window slice = collapsed.first_slice_window_2D(); + Window collapsed = window.collapse_if_possible(_full_window, Window::DimZ); + Window slice = collapsed.first_slice_window_3D(); do { unsigned int idx = 0; - add_2D_tensor_argument(idx, _tensor, slice); + add_3D_tensor_argument(idx, _tensor, slice); enqueue(queue, *this, slice); } - while(collapsed.slide_window_slice_2D(slice)); + while(collapsed.slide_window_slice_3D(slice)); } } // namespace arm_compute -- cgit v1.2.1