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/CLKernelLibrary.cpp | 5 + src/core/CL/cl_kernels/copy_tensor.cl | 17 +- src/core/CL/cl_kernels/crop_tensor.cl | 96 +++++++++++ src/core/CL/cl_kernels/memset.cl | 15 +- src/core/CL/kernels/CLCopyKernel.cpp | 122 ++++++++++---- src/core/CL/kernels/CLCropKernel.cpp | 132 +++++++++++++++ src/core/CL/kernels/CLMemsetKernel.cpp | 47 ++++-- src/runtime/CL/functions/CLCropResize.cpp | 272 ++++++++++++++++++++++++++++++ 8 files changed, 652 insertions(+), 54 deletions(-) create mode 100644 src/core/CL/cl_kernels/crop_tensor.cl create mode 100644 src/core/CL/kernels/CLCropKernel.cpp create mode 100644 src/runtime/CL/functions/CLCropResize.cpp (limited to 'src') diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp index b7973efa9d..5af8a09723 100644 --- a/src/core/CL/CLKernelLibrary.cpp +++ b/src/core/CL/CLKernelLibrary.cpp @@ -212,6 +212,7 @@ const std::map CLKernelLibrary::_kernel_program_map = { "copy_plane", "channel_extract.cl" }, { "copy_planes_3p", "channel_combine.cl" }, { "copy_to_keypoint", "fast_corners.cl" }, + { "crop_tensor", "crop_tensor.cl" }, { "deconvolution_upsample", "deconvolution_layer.cl" }, { "depthwise_convolution_3x3", "depthwise_convolution.cl" }, { "depthwise_convolution_3x3_f16", "depthwise_convolution.cl" }, @@ -607,6 +608,10 @@ const std::map CLKernelLibrary::_program_source_map = { "copy_tensor.cl", #include "./cl_kernels/copy_tensor.clembed" + }, + { + "crop_tensor.cl", +#include "./cl_kernels/crop_tensor.clembed" }, { "upsample_layer.cl", diff --git a/src/core/CL/cl_kernels/copy_tensor.cl b/src/core/CL/cl_kernels/copy_tensor.cl index 4bbbf11bea..f4366b889a 100644 --- a/src/core/CL/cl_kernels/copy_tensor.cl +++ b/src/core/CL/cl_kernels/copy_tensor.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -77,6 +77,7 @@ __kernel void copy_pad_tensor( } #endif // Compile time constants +#if defined(DATA_TYPE) /** Performs a copy of input tensor to the output tensor. * * @param[in] in_ptr Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 @@ -103,6 +104,16 @@ __kernel void copy_tensor( Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(in); Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out); +#if defined(VEC_SIZE) + +#if defined(LAST_ACCESSED_X) + // Check if access on width gets out of bounds + // If it does then shift access vector to access elements within bounds + const int shift = max((int)(get_global_id(0) * VEC_SIZE) - (int)LAST_ACCESSED_X, 0); + in.ptr -= shift * in.stride_x; + out.ptr -= shift * out.stride_x; +#endif // defined(LAST_ACCESSED_X) + // Load data VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in.ptr); @@ -110,4 +121,8 @@ __kernel void copy_tensor( // Store result VSTORE(VEC_SIZE) (data, 0, (__global DATA_TYPE *)out.ptr); +#else // defined(VEC_SIZE) + *((__global DATA_TYPE *)(out.ptr)) = *((__global DATA_TYPE *)(in.ptr)); +#endif // defined(VEC_SIZE) } +#endif // defined(DATA_TYPE) \ No newline at end of file diff --git a/src/core/CL/cl_kernels/crop_tensor.cl b/src/core/CL/cl_kernels/crop_tensor.cl new file mode 100644 index 0000000000..55f8544a10 --- /dev/null +++ b/src/core/CL/cl_kernels/crop_tensor.cl @@ -0,0 +1,96 @@ +/* + * Copyright (c) 2019 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "helpers.h" + +#if defined(DATA_TYPE) // Compile time constants + +/** Performs a copy of input tensor to the output tensor. + * + * @param[in] in_ptr Pointer to the source tensor. Supported data types: U16/S16/F16/U32/S32/F32 + * @param[in] in_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] in_step_x input_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] in_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] in_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] in_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] in_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] in_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] out_ptr Pointer to the destination tensor. Supported data types: same as @p in_ptr + * @param[in] out_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] out_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] out_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] out_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] out_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] out_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] in_offset_y The initial offset of the input address along Y. + * @param[in] in_offset_z The initial offset of the input address along Z. + */ +__kernel void crop_tensor( + TENSOR3D_DECLARATION(in), + TENSOR3D_DECLARATION(out), + int in_offset_y, + int in_offset_z) +{ + Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(in); + Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out); + + const int in_x = get_global_id(0) * (in_step_x / in_stride_x); + +#if defined(WIDTH_FLIPPED) + const int in_y = in_offset_y - get_global_id(1); +#else // defined(WIDTH_FLIPPED) + const int in_y = in_offset_y + get_global_id(1); +#endif // defined(WIDTH_FLIPPED) + +#if defined(HEIGHT_FLIPPED) + const int in_z = in_offset_z - get_global_id(2); +#else // defined(HEIGHT_FLIPPED) + const int in_z = in_offset_z + get_global_id(2); +#endif // defined(HEIGHT_FLIPPED) + +#if defined(VEC_SIZE) + +#if defined(LAST_ACCESSED_X) + // Check if access on width gets out of bounds + // If it does then shift access vector to access elements within bounds + const int shift = max((int)(get_global_id(0) * VEC_SIZE) - (int)LAST_ACCESSED_X, 0); + in.ptr -= shift * in.stride_x; + out.ptr -= shift * out.stride_x; +#endif // defined(LAST_ACCESSED_X) + + __global const uchar *input_addr = tensor3D_offset(&in, in_x, in_y, in_z); + + // Load data + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr); + + // Store result + VSTORE(VEC_SIZE) + (CONVERT(data, VEC_DATA_TYPE(float, VEC_SIZE)), 0, (__global float *)out.ptr); +#else // defined(VEC_SIZE) + *((__global float *)(out.ptr)) = CONVERT(*((__global DATA_TYPE *)tensor3D_offset(&in, in_x, in_y, in_z)), float); +#endif // defined(VEC_SIZE) +} + +#endif // defined(DATA_TYPE) && defined(LAST_ACCESSED_X) \ No newline at end of file diff --git a/src/core/CL/cl_kernels/memset.cl b/src/core/CL/cl_kernels/memset.cl index 80b34ebdf4..7d8e0ef53f 100644 --- a/src/core/CL/cl_kernels/memset.cl +++ b/src/core/CL/cl_kernels/memset.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -41,24 +41,27 @@ * @param[in] value The value used to fill the pages of the tensor */ __kernel void memset( - IMAGE_DECLARATION(tensor)) + TENSOR3D_DECLARATION(tensor)) { - Image tensor = CONVERT_TO_IMAGE_STRUCT(tensor); + Tensor3D tensor = CONVERT_TO_TENSOR3D_STRUCT(tensor); -#if defined(VEC_SIZE) && defined(LAST_ACCESSED_X) +#if defined(VEC_SIZE) + +#if defined(LAST_ACCESSED_X) // Check if access on width gets out of bounds // If it does shift access vector to access elements within bounds const int xi = (int)(get_global_id(0) * VEC_SIZE); tensor.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * tensor_stride_x; +#endif // defined(LAST_ACCESSED_X) VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) data = (DATA_TYPE)(CONSTANT_VALUE); VSTORE(VEC_SIZE) (data, 0, (__global DATA_TYPE *)tensor.ptr); -#else // !defined(VEC_SIZE) || !defined(LAST_ACCESSED_X) +#else // !defined(VEC_SIZE) *((__global DATA_TYPE *)(tensor.ptr)) = (DATA_TYPE)(CONSTANT_VALUE); -#endif // defined(VEC_SIZE) && defined(LAST_ACCESSED_X) +#endif // defined(VEC_SIZE) } #endif // Check for compile time constants diff --git a/src/core/CL/kernels/CLCopyKernel.cpp b/src/core/CL/kernels/CLCopyKernel.cpp index e14e5dafab..30a0b8fcb3 100644 --- a/src/core/CL/kernels/CLCopyKernel.cpp +++ b/src/core/CL/kernels/CLCopyKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -37,38 +37,58 @@ namespace arm_compute { namespace { -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding = PaddingList()) +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding = PaddingList(), Window *output_window = nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_ON(!padding.empty() && output_window != nullptr); ARM_COMPUTE_RETURN_ERROR_ON(padding.size() > 4); // Validate output if initialized if(output->total_size() != 0) { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding), output->tensor_shape()); + if(output_window == nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding), output->tensor_shape()); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(input->tensor_shape(), output_window->shape()); + } ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); } return Status{}; } -std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *output) +std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, Window *output_window) { // Output auto inizialitation if not yet initialized auto_init_if_empty(*output, *input); // Configure window - const unsigned int num_elems_processed_per_iteration = 16 / input->element_size(); + const unsigned int vec_size_x = 16 / input->element_size(); - Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); + if(output_window == nullptr) + { + // Create and update the window (if needed) + Window win = calculate_max_window(*input, Steps(vec_size_x)); - AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); - AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); + AccessWindowHorizontal input_access(input, 0, vec_size_x); + AccessWindowHorizontal output_access(output, 0, vec_size_x); - bool window_changed = update_window_and_padding(win, input_access, output_access); + bool window_changed = update_window_and_padding(win, input_access, output_access); - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_pair(err, win); + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); + } + else + { + Window win = calculate_max_window(*input); + return std::make_pair(Status{}, win); + } } std::pair validate_and_configure_window_with_padding(ITensorInfo *input, ITensorInfo *output, const PaddingList &padding) @@ -131,14 +151,14 @@ void add_padding_as_build_options(const PaddingList &padding, CLBuildOptions &bu } // namespace CLCopyKernel::CLCopyKernel() - : _input(nullptr), _output(nullptr) + : _input(nullptr), _output(nullptr), _output_window(), _has_output_window(false) { } -void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output, const PaddingList &padding) +void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output, const PaddingList &padding, Window *output_window) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), padding)); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), padding, output_window)); _input = input; _output = output; @@ -147,21 +167,44 @@ void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output, const Pa CLBuildOptions build_opts; build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); - const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size(); - build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); - std::pair win_config; + const unsigned int vec_size_x = 16 / input->info()->element_size(); + if(padding.empty()) { + // Configure window + win_config = validate_and_configure_window(input->info(), output->info(), output_window); + + if(output_window != nullptr) + { + _has_output_window = true; + _output_window = Window(*output_window); + const int width_x = output_window->num_iterations(0); + const bool multi_access_x = width_x >= static_cast(vec_size_x); + const bool remainder_x = width_x % vec_size_x > 0; + + if(multi_access_x) + { + _output_window.set(Window::DimX, Window::Dimension(output_window->x().start(), ceil_to_multiple(output_window->x().end(), vec_size_x), vec_size_x)); + win_config.second.set(Window::DimX, Window::Dimension(win_config.second.x().start(), ceil_to_multiple(win_config.second.x().end(), vec_size_x), vec_size_x)); + } + + 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 && remainder_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max(width_x - vec_size_x, 0))); + } + else + { + build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); + } + // Build kernel _kernel = static_cast(CLKernelLibrary::get().create_kernel("copy_tensor", build_opts.options())); - - // Configure window - win_config = validate_and_configure_window(input->info(), output->info()); } else { + build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); + // Add compile time options add_padding_as_build_options(padding, build_opts); @@ -185,13 +228,13 @@ void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output, const Pa ICLKernel::configure_internal(win_config.second); } -Status CLCopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output, const PaddingList &padding) +Status CLCopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output, const PaddingList &padding, Window *output_window) { - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, padding)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, padding, output_window)); if(padding.empty()) { - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), output_window).first); } else { @@ -206,16 +249,33 @@ void CLCopyKernel::run(const Window &window, cl::CommandQueue &queue) ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); - Window slice = collapsed.first_slice_window_3D(); + Window slice; - do + if(_has_output_window) + { + slice = window.first_slice_window_3D(); + Window out_slice = _output_window.first_slice_window_3D(); + do + { + unsigned int idx = 0; + add_3D_tensor_argument(idx, _input, slice); + add_3D_tensor_argument(idx, _output, out_slice); + enqueue(queue, *this, slice); + } + while(window.slide_window_slice_3D(slice) && _output_window.slide_window_slice_3D(out_slice)); + } + else { - unsigned int idx = 0; - add_3D_tensor_argument(idx, _input, slice); - add_3D_tensor_argument(idx, _output, slice); - enqueue(queue, *this, slice); + Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); + slice = collapsed.first_slice_window_3D(); + do + { + unsigned int idx = 0; + add_3D_tensor_argument(idx, _input, slice); + add_3D_tensor_argument(idx, _output, slice); + enqueue(queue, *this, slice); + } + while(collapsed.slide_window_slice_3D(slice)); } - while(collapsed.slide_window_slice_3D(slice)); } } // namespace arm_compute diff --git a/src/core/CL/kernels/CLCropKernel.cpp b/src/core/CL/kernels/CLCropKernel.cpp new file mode 100644 index 0000000000..f8a2456d4a --- /dev/null +++ b/src/core/CL/kernels/CLCropKernel.cpp @@ -0,0 +1,132 @@ +/* + * Copyright (c) 2019 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/CL/kernels/CLCropKernel.h" + +#include "arm_compute/core/CL/CLHelpers.h" +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/CL/ICLTensor.h" +#include "arm_compute/core/CPP/Validate.h" +#include "arm_compute/core/IAccessWindow.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Window.h" + +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/helpers/bit_ops.h" +#include "arm_compute/core/utils/helpers/tensor_transform.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" + +#include + +namespace arm_compute +{ +CLCropKernel::CLCropKernel() + : _input(nullptr), _output(nullptr), _start(), _batch_index(0), _extrapolation_value(0) +{ +} + +void CLCropKernel::configure(const ICLTensor *input, ICLTensor *output, Coordinates2D start, Coordinates2D end, uint32_t batch_index, float extrapolation_value, Window *output_window) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_THROW_ON(validate(input->info(), output->info(), start, end, batch_index, extrapolation_value, output_window)); + + _input = input; + _output = output; + _start = start; + _batch_index = batch_index; + _extrapolation_value = extrapolation_value; + + const int vec_size_x = 4; + // Create and update the window (if needed) + Window win = calculate_max_window(*output->info()); + + if(output_window != nullptr) + { + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(win, *output_window); + win = *output_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)); + } + ICLKernel::configure_internal(win); + + // Create kernel + CLBuildOptions build_opts; + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->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 && remainder_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max(output_width_x - vec_size_x, 0))); + build_opts.add_option_if(start.x > end.x, "-DWIDTH_FLIPPED="); + build_opts.add_option_if(start.y > end.y, "-DHEIGHT_FLIPPED="); + _kernel = static_cast(CLKernelLibrary::get().create_kernel("crop_tensor", build_opts.options())); +} + +Status CLCropKernel::validate(const ITensorInfo *input, const ITensorInfo *output, Coordinates2D start, Coordinates2D end, uint32_t batch_index, float extrapolation_value, Window *output_window) +{ + ARM_COMPUTE_UNUSED(extrapolation_value, output_window); + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U16, DataType::S16, DataType::F16, DataType::U32, DataType::S32, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC); + ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape().num_dimensions() > 4); + ARM_COMPUTE_RETURN_ERROR_ON(start.x < 0 || start.y < 0 || end.x < 0 || end.y < 0); + ARM_COMPUTE_RETURN_ERROR_ON(start.x >= static_cast(input->dimension(1)) || start.y >= static_cast(input->dimension(2)) + || end.x >= static_cast(input->dimension(1)) || end.y >= static_cast(input->dimension(2))); + ARM_COMPUTE_RETURN_ERROR_ON(batch_index >= input->dimension(3)); + if(output_window != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON(output_window->x().step() != 1); + } + if(output->total_size() > 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(output, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); + ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 3); + } + return Status{}; +} + +void CLCropKernel::run(const Window &window, cl::CommandQueue &queue) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); + + Window in_slice = Window(); + in_slice.use_tensor_dimensions(_input->info()->tensor_shape()); + in_slice.set(Window::DimX, Window::Dimension(in_slice.x().start(), ceil_to_multiple(in_slice.x().end(), window.x().step()), window.x().step())); + in_slice.set(3, Window::Dimension(_batch_index, _batch_index + 1, 1)); + + unsigned int idx = 0; + add_3D_tensor_argument(idx, _input, in_slice); + add_3D_tensor_argument(idx, _output, window); + add_argument(idx, _start.x); + add_argument(idx, _start.y); + enqueue(queue, *this, window); +} +} // namespace arm_compute 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 diff --git a/src/runtime/CL/functions/CLCropResize.cpp b/src/runtime/CL/functions/CLCropResize.cpp new file mode 100644 index 0000000000..2cacef1bb1 --- /dev/null +++ b/src/runtime/CL/functions/CLCropResize.cpp @@ -0,0 +1,272 @@ +/* + * Copyright (c) 2019 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/core/CL/CLHelpers.h" + +#include "arm_compute/runtime/CL/CLScheduler.h" +#include "arm_compute/runtime/CL/functions/CLCropResize.h" + +#include + +namespace arm_compute +{ +namespace +{ +inline void configure_crop(const ICLTensor *input, ICLTensor *crop_boxes, ICLTensor *box_ind, ICLTensor *output, uint32_t crop_box_ind, Coordinates &start, Coordinates &end, uint32_t &batch_index) +{ + batch_index = *(reinterpret_cast(box_ind->ptr_to_element(Coordinates(crop_box_ind)))); + + // _crop_box_ind is used to index crop_boxes and retrieve the appropriate crop box. + // The crop box is specified by normalized coordinates [y0, x0, y1, x1]. + const float x0 = *reinterpret_cast(crop_boxes->ptr_to_element(Coordinates(1, crop_box_ind))); + const float y0 = *reinterpret_cast(crop_boxes->ptr_to_element(Coordinates(0, crop_box_ind))); + const float x1 = *reinterpret_cast(crop_boxes->ptr_to_element(Coordinates(3, crop_box_ind))); + const float y1 = *reinterpret_cast(crop_boxes->ptr_to_element(Coordinates(2, crop_box_ind))); + // The normalized coordinates are scaled to retrieve the floating point image coordinates which are rounded to integers. + start = Coordinates(std::floor(x0 * (input->info()->tensor_shape()[1] - 1) + 0.5f), + std::floor(y0 * (input->info()->tensor_shape()[2] - 1) + 0.5f)); + end = Coordinates(std::floor(x1 * (input->info()->tensor_shape()[1] - 1) + 0.5f), + std::floor(y1 * (input->info()->tensor_shape()[2] - 1) + 0.5f)); + const TensorShape out_shape(input->info()->tensor_shape()[0], abs(end[0] - start[0]) + 1, abs(end[1] - start[1]) + 1); + output->info()->set_tensor_shape(out_shape); +} + +inline void run_crop(const ICLTensor *input, ICLTensor *output, uint32_t batch_index, Coordinates start, Coordinates end, float extrapolation_value) +{ + bool is_width_flipped = end[0] < start[0]; + bool is_height_flipped = end[1] < start[1]; + /** The number of rows out of bounds at the start and end of output. */ + int32_t rows_out_of_bounds[2]; + /** The number of columns out of bounds at the start and end of output. */ + int32_t cols_out_of_bounds[2]; + if(is_height_flipped) + { + rows_out_of_bounds[0] = start[1] >= static_cast(input->info()->dimension(2)) ? std::min(start[1] - input->info()->dimension(2) + 1, output->info()->dimension(2)) : 0; + rows_out_of_bounds[1] = end[1] < 0 ? std::min(-end[1], static_cast(output->info()->dimension(2))) : 0; + } + else + { + rows_out_of_bounds[0] = start[1] < 0 ? std::min(-start[1], static_cast(output->info()->dimension(2))) : 0; + rows_out_of_bounds[1] = end[1] >= static_cast(input->info()->dimension(2)) ? std::min(end[1] - input->info()->dimension(2) + 1, output->info()->dimension(2)) : 0; + } + if(is_width_flipped) + { + cols_out_of_bounds[0] = start[0] >= static_cast(input->info()->dimension(1)) ? std::min(start[0] - input->info()->dimension(1) + 1, output->info()->dimension(1)) : 0; + cols_out_of_bounds[1] = end[0] < 0 ? std::min(-end[0], static_cast(output->info()->dimension(1))) : 0; + } + else + { + cols_out_of_bounds[0] = start[0] < 0 ? std::min(-start[0], static_cast(output->info()->dimension(1))) : 0; + cols_out_of_bounds[1] = end[0] >= static_cast(input->info()->dimension(1)) ? std::min(end[0] - input->info()->dimension(1) + 1, output->info()->dimension(1)) : 0; + } + + Window full_window = calculate_max_window(*output->info()); + + // Full output window: + // -------------------------------- + // | Out of bounds | + // | rows before | + // |------------------------------| + // | Out of | In | Out of | + // | bounds | bounds | bounds | + // | cols | elements | cols | + // | before | copied | after | + // | | from input | | + // |------------------------------| + // | Out of bounds | + // | rows after | + // |------------------------------| + // Use a separate output window for each section of the full output window. + // Fill all output rows that have no elements that are within the input bounds + // with the extrapolation value using memset. + // First for the rows before the in bounds rows. + if(rows_out_of_bounds[0] > 0) + { + Window slice_fill_rows_before(full_window); + slice_fill_rows_before.set(2, Window::Dimension(0, rows_out_of_bounds[0], 1)); + auto kernel = arm_compute::support::cpp14::make_unique(); + kernel->configure(output, extrapolation_value, &slice_fill_rows_before); + CLScheduler::get().enqueue(*kernel); + } + + Window slice_in(full_window); + slice_in.set(2, Window::Dimension(rows_out_of_bounds[0], output->info()->dimension(2) - rows_out_of_bounds[1], 1)); + slice_in.set(1, Window::Dimension(cols_out_of_bounds[0], output->info()->dimension(1) - cols_out_of_bounds[1], 1)); + + int rows_in_bounds = static_cast(output->info()->dimension(2)) - rows_out_of_bounds[0] - rows_out_of_bounds[1]; + if(rows_in_bounds > 0) + { + // Fill all elements that share a row with an in bounds element with the extrapolation value. + if(cols_out_of_bounds[0] > 0) + { + Window slice_fill_cols_before(slice_in); + slice_fill_cols_before.set(1, Window::Dimension(0, cols_out_of_bounds[0], 1)); + auto kernel = arm_compute::support::cpp14::make_unique(); + kernel->configure(output, extrapolation_value, &slice_fill_cols_before); + CLScheduler::get().enqueue(*kernel); + } + + if(cols_out_of_bounds[1] > 0) + { + Window slice_fill_cols_after(slice_in); + slice_fill_cols_after.set(1, Window::Dimension(output->info()->dimension(1) - cols_out_of_bounds[1], output->info()->dimension(1), 1)); + auto kernel = arm_compute::support::cpp14::make_unique(); + kernel->configure(output, extrapolation_value, &slice_fill_cols_after); + CLScheduler::get().enqueue(*kernel); + } + + // Copy all elements within the input bounds from the input tensor. + int cols_in_bounds = static_cast(output->info()->dimension(1)) - cols_out_of_bounds[0] - cols_out_of_bounds[1]; + if(cols_in_bounds > 0) + { + Coordinates2D start_in{ is_width_flipped ? start[0] - cols_out_of_bounds[0] : start[0] + cols_out_of_bounds[0], + is_height_flipped ? start[1] - rows_out_of_bounds[0] : start[1] + rows_out_of_bounds[0] }; + Coordinates2D end_in{ is_width_flipped ? start_in.x - cols_in_bounds + 1 : start_in.x + cols_in_bounds - 1, + is_height_flipped ? start_in.y - rows_in_bounds + 1 : start_in.y + rows_in_bounds - 1 }; + auto kernel = arm_compute::support::cpp14::make_unique(); + + kernel->configure(input, output, start_in, end_in, batch_index, extrapolation_value, &slice_in); + CLScheduler::get().enqueue(*kernel); + } + } + + // Fill all rows after the in bounds elements with the extrapolation value. + if(rows_out_of_bounds[1] > 0) + { + Window slice_fill_rows_after(full_window); + slice_fill_rows_after.set(2, Window::Dimension(output->info()->dimension(2) - rows_out_of_bounds[1], output->info()->dimension(2), 1)); + auto kernel = arm_compute::support::cpp14::make_unique(); + kernel->configure(output, extrapolation_value, &slice_fill_rows_after); + CLScheduler::get().enqueue(*kernel); + } +} +} // namespace + +CLCropResize::CLCropResize() + : _input(nullptr), _boxes(nullptr), _box_ind(nullptr), _output(nullptr), _num_boxes(0), _method(), _extrapolation_value(0), _scale(), _copy() +{ +} + +Status CLCropResize::validate(const ITensorInfo *input, ITensorInfo *boxes, ITensorInfo *box_ind, const ITensorInfo *output, + Coordinates2D crop_size, InterpolationPolicy method, float extrapolation_value) +{ + ARM_COMPUTE_RETURN_ERROR_ON(crop_size.x <= 0 || crop_size.y <= 0); + ARM_COMPUTE_RETURN_ERROR_ON(method == InterpolationPolicy::AREA); + ARM_COMPUTE_RETURN_ERROR_ON(boxes->tensor_shape()[0] != 4); + ARM_COMPUTE_RETURN_ERROR_ON(boxes->tensor_shape()[1] != box_ind->tensor_shape()[0]); + TensorInfo temp_info; + ARM_COMPUTE_RETURN_ON_ERROR(CLCropKernel::validate(input->clone().get(), &temp_info, { 0, 0 }, { 1, 1 }, input->dimension(3) - 1, extrapolation_value)); + if(output->total_size() > 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(output, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); + TensorShape out_shape(input->tensor_shape()[0], crop_size.x, crop_size.y, boxes->tensor_shape()[1]); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), out_shape); + } + return Status{}; +} + +void CLCropResize::configure(const ICLTensor *input, ICLTensor *boxes, ICLTensor *box_ind, ICLTensor *output, Coordinates2D crop_size, + InterpolationPolicy method, float extrapolation_value) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_THROW_ON(CLCropResize::validate(input->info(), boxes->info(), box_ind->info(), output->info(), crop_size, method, extrapolation_value)); + + _num_boxes = boxes->info()->tensor_shape()[1]; + TensorShape out_shape(input->info()->tensor_shape()[0], crop_size.x, crop_size.y); + + _input = input; + _boxes = boxes; + _box_ind = box_ind; + _output = output; + _method = method; + _extrapolation_value = extrapolation_value; + + // For each crop box: + // - The initial cropped image is produced as specified by boxes[i] from the 3D image input[box_ind[i]]. + // Possibly using a CLCropKernel and up to four CLMemsetKernels. + // - A tensor is required to hold this initial cropped image. + // - A scale function is used to resize the cropped image to the size specified by crop_size. + // - A tensor is required to hold the final scaled image before it is copied into the 4D output + // that will hold all final cropped and scaled 3D images using CLCopyKernel. + _crop_results = arm_compute::support::cpp14::make_unique(_num_boxes); + _scale = arm_compute::support::cpp14::make_unique(_num_boxes); + _scaled_results = arm_compute::support::cpp14::make_unique(_num_boxes); + _copy = arm_compute::support::cpp14::make_unique(_num_boxes); + + for(unsigned int i = 0; i < _num_boxes; ++i) + { + TensorInfo crop_result_info(1, DataType::F32); + crop_result_info.set_data_layout(DataLayout::NHWC); + _crop_results[i].allocator()->init(crop_result_info); + + TensorInfo scaled_result_info(out_shape, 1, DataType::F32); + scaled_result_info.set_data_layout(DataLayout::NHWC); + _scaled_results[i].allocator()->init(scaled_result_info); + } +} + +void CLCropResize::run() +{ + ARM_COMPUTE_ERROR_ON_MSG(_output == nullptr, "Unconfigured function"); + // The contents of _boxes and _box_ind are required to calculate the shape + // of the initial cropped image and thus are required to configure the + // kernels used for cropping and scaling. + _boxes->map(CLScheduler::get().queue()); + _box_ind->map(CLScheduler::get().queue()); + for(unsigned int i = 0; i < _num_boxes; ++i) + { + // Size of the crop box in _boxes and thus the shape of _crop_results[i] + // may not be known until run-time and so the kernels cannot be configured until then. + uint32_t batch_index; + Coordinates start, end; + configure_crop(_input, _boxes, _box_ind, &_crop_results[i], i, start, end, batch_index); + _scale[i].configure(&_crop_results[i], &_scaled_results[i], _method, BorderMode::CONSTANT, PixelValue(_extrapolation_value), SamplingPolicy::TOP_LEFT); + + Window win = calculate_max_window(*_output->info()); + win.set(3, Window::Dimension(i, i + 1, 1)); + _copy[i].configure(&_scaled_results[i], _output, PaddingList(), &win); + + _crop_results[i].allocator()->allocate(); + _scaled_results[i].allocator()->allocate(); + + run_crop(_input, &_crop_results[i], batch_index, start, end, _extrapolation_value); + } + _boxes->unmap(CLScheduler::get().queue()); + _box_ind->unmap(CLScheduler::get().queue()); + CLScheduler::get().sync(); + for(unsigned int i = 0; i < _num_boxes; ++i) + { + // Scale the cropped image + _scale[i].run(); + } + CLScheduler::get().sync(); + for(unsigned int i = 0; i < _num_boxes; ++i) + { + // Copy scaled image into output. + CLScheduler::get().enqueue(_copy[i]); + } + CLScheduler::get().sync(); +} +} // namespace arm_compute \ No newline at end of file -- cgit v1.2.1