diff options
Diffstat (limited to 'src/core')
-rw-r--r-- | src/core/CL/CLKernels.h | 3 | ||||
-rw-r--r-- | src/core/CL/cl_kernels/nchw/remap.cl | 133 | ||||
-rw-r--r-- | src/core/CL/cl_kernels/nhwc/remap.cl | 180 | ||||
-rw-r--r-- | src/core/CL/kernels/CLRemapKernel.cpp | 183 | ||||
-rw-r--r-- | src/core/CL/kernels/CLRemapKernel.h | 88 | ||||
-rw-r--r-- | src/core/NEON/NEKernels.h | 3 | ||||
-rw-r--r-- | src/core/NEON/kernels/NERemapKernel.cpp | 326 | ||||
-rw-r--r-- | src/core/NEON/kernels/NERemapKernel.h | 86 |
8 files changed, 2 insertions, 1000 deletions
diff --git a/src/core/CL/CLKernels.h b/src/core/CL/CLKernels.h index 0c295aae6a..63be7b1ea8 100644 --- a/src/core/CL/CLKernels.h +++ b/src/core/CL/CLKernels.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2021 Arm Limited. + * Copyright (c) 2016-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -56,7 +56,6 @@ #include "src/core/CL/kernels/CLROIPoolingLayerKernel.h" #include "src/core/CL/kernels/CLRangeKernel.h" #include "src/core/CL/kernels/CLReductionOperationKernel.h" -#include "src/core/CL/kernels/CLRemapKernel.h" #include "src/core/CL/kernels/CLReorgLayerKernel.h" #include "src/core/CL/kernels/CLReverseKernel.h" #include "src/core/CL/kernels/CLSelectKernel.h" diff --git a/src/core/CL/cl_kernels/nchw/remap.cl b/src/core/CL/cl_kernels/nchw/remap.cl deleted file mode 100644 index fab88a1682..0000000000 --- a/src/core/CL/cl_kernels/nchw/remap.cl +++ /dev/null @@ -1,133 +0,0 @@ -/* - * Copyright (c) 2017-2021 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" -#include "warp_helpers.h" - -#ifndef DEPTH_OUT -/** Performs a remapping of an input image to an output given two remapping image using nearest neighbor as interpolation. - * - * This kernel performs remapping with this method of pixel coordinate translation: - * out(x,y) = in(mapx(x,y), mapy(x,y)); - * - * @param[in] in_ptr Pointer to the source image. Supported data types: U8. - * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] in_step_x in_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] in_step_y in_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] in_offset_first_element_in_bytes Offset of the first element in the source image - * @param[out] out_ptr Pointer to the destination image. Supported data types: U8. - * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) - * @param[in] out_step_x out_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) - * @param[in] out_step_y out_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] out_offset_first_element_in_bytes Offset of the first element in the destination image - * @param[in] mapx_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapx_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapx_step_x mapx_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapx_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapx_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapx_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] mapy_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapy_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapy_step_x mapy_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapy_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapy_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapy_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] width Width of the input image - * @param[in] height Height of the input image - */ -__kernel void remap_nearest_neighbour_nchw( - IMAGE_DECLARATION(in), - IMAGE_DECLARATION(out), - IMAGE_DECLARATION(mapx), - IMAGE_DECLARATION(mapy), - const float width, - const float height) -{ - Image in = CONVERT_TO_IMAGE_STRUCT_NO_STEP(in); - Image out = CONVERT_TO_IMAGE_STRUCT(out); - Image mapx = CONVERT_TO_IMAGE_STRUCT(mapx); - Image mapy = CONVERT_TO_IMAGE_STRUCT(mapy); - - float4 mapx_coords = vload4(0, (__global float *)mapx.ptr); - float4 mapy_coords = vload4(0, (__global float *)mapy.ptr); - float8 map_coords = (float8)(mapx_coords.s0, mapy_coords.s0, mapx_coords.s1, mapy_coords.s1, - mapx_coords.s2, mapy_coords.s2, mapx_coords.s3, mapy_coords.s3); - - vstore4(read_texels4(&in, convert_int8(clamp_to_border(map_coords, width, height))), 0, out.ptr); -} - -/** Performs a remapping of an input image to an output given two remapping image using bilinear as interpolation. - * - * This kernel performs remapping with this method of pixel coordinate translation: - * out(x,y) = in(mapx(x,y), mapy(x,y)); - * - * @param[in] in_ptr Pointer to the source image. Supported data types: U8. - * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] in_step_x in_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] in_step_y in_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] in_offset_first_element_in_bytes Offset of the first element in the source image - * @param[out] out_ptr Pointer to the destination image. Supported data types: U8. - * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) - * @param[in] out_step_x out_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) - * @param[in] out_step_y out_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] out_offset_first_element_in_bytes Offset of the first element in the destination image - * @param[in] mapx_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapx_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapx_step_x mapx_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapx_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapx_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapx_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] mapy_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapy_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapy_step_x mapy_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapy_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapy_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapy_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] width Width of the input image - * @param[in] height Height of the input image - */ -__kernel void remap_bilinear_nchw( - IMAGE_DECLARATION(in), - IMAGE_DECLARATION(out), - IMAGE_DECLARATION(mapx), - IMAGE_DECLARATION(mapy), - const float width, - const float height) -{ - Image in = CONVERT_TO_IMAGE_STRUCT_NO_STEP(in); - Image out = CONVERT_TO_IMAGE_STRUCT(out); - Image mapx = CONVERT_TO_IMAGE_STRUCT(mapx); - Image mapy = CONVERT_TO_IMAGE_STRUCT(mapy); - - float4 mapx_coords = vload4(0, (__global float *)mapx.ptr); - float4 mapy_coords = vload4(0, (__global float *)mapy.ptr); - float8 map_coords = (float8)(mapx_coords.s0, mapy_coords.s0, mapx_coords.s1, mapy_coords.s1, - mapx_coords.s2, mapy_coords.s2, mapx_coords.s3, mapy_coords.s3); - - vstore4(bilinear_interpolate(&in, clamp_to_border(map_coords, width, height), width, height), 0, out.ptr); -} -#endif // DEPTH_OUT
\ No newline at end of file diff --git a/src/core/CL/cl_kernels/nhwc/remap.cl b/src/core/CL/cl_kernels/nhwc/remap.cl deleted file mode 100644 index 0b629fe6c9..0000000000 --- a/src/core/CL/cl_kernels/nhwc/remap.cl +++ /dev/null @@ -1,180 +0,0 @@ -/* - * Copyright (c) 2017-2021 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" -#include "warp_helpers.h" - -#ifdef DEPTH_OUT -/** Performs a remapping of an input image to an output given two remapping image using nearest neighbor as interpolation. - * Also applies constant border value, "border_val", if "CONSTANT_BORDER" is set. - * - * This kernel performs remapping with this method of pixel coordinate translation: - * out(x,y) = in(mapx(x,y), mapy(x,y)); - * - * @param[in] in_ptr Pointer to the source image. Supported data types: U8,F16. - * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] in_step_x in_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] in_step_y in_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] in_offset_first_element_in_bytes Offset of the first element in the source image - * @param[out] out_ptr Pointer to the destination image. Supported data types: U8,F16. - * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) - * @param[in] out_step_x out_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) - * @param[in] out_step_y out_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] out_offset_first_element_in_bytes Offset of the first element in the destination image - * @param[in] mapx_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapx_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapx_step_x mapx_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapx_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapx_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapx_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] mapy_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapy_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapy_step_x mapy_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapy_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapy_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapy_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] width Width of the input image - * @param[in] height Height of the input image - * @param[in] border_val Value to use for border around input tensor when in CONSTANT border is selected - */ -__kernel void remap_nearest_neighbour_nhwc( - TENSOR4D_DECLARATION(in), - TENSOR4D_DECLARATION(out), - TENSOR4D_DECLARATION(mapx), - TENSOR4D_DECLARATION(mapy), - const float width, - const float height -#ifdef CONSTANT_BORDER - , - const DATA_TYPE border_val -#endif // CONSTANT_BORDER -) -{ - Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(in, 0); - Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT(out, DEPTH_OUT); - Tensor4D mapx = CONVERT_TO_TENSOR4D_STRUCT(mapx, DEPTH_OUT); - Tensor4D mapy = CONVERT_TO_TENSOR4D_STRUCT(mapy, DEPTH_OUT); - - float mapx_coord = (float) * (__global float *)mapx.ptr; - float mapy_coord = (float) * (__global float *)mapy.ptr; - -#ifdef CONSTANT_BORDER - if(mapx_coord < 0 || mapx_coord > width - 1 || mapy_coord < 0 || mapy_coord > height - 1) - { - *((__global DATA_TYPE *)out.ptr) = border_val; - return; - } -#else // CONSTANT_BORDER - mapx_coord = clamp(mapx_coord, 0.0f, width - 1); - mapy_coord = clamp(mapy_coord, 0.0f, height - 1); -#endif // CONSTANT_BORDER - *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(mapx_coord), convert_int(mapy_coord), (get_global_id(2) / DEPTH_OUT))); -} - -/** Performs a remapping of an input image to an output given two remapping image using bilinear as interpolation. - * Also applies constant border value, "border_val", if "CONSTANT_BORDER" is set. - * - * This kernel performs remapping with this method of pixel coordinate translation: - * out(x,y) = in(mapx(x,y), mapy(x,y)); - * - * @param[in] in_ptr Pointer to the source image. Supported data types: U8,F16. - * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] in_step_x in_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] in_step_y in_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] in_offset_first_element_in_bytes Offset of the first element in the source image - * @param[out] out_ptr Pointer to the destination image. Supported data types: U8,F16. - * @param[in] out_stride_x Stride of the destination image in X dimension (in bytes) - * @param[in] out_step_x out_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes) - * @param[in] out_step_y out_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] out_offset_first_element_in_bytes Offset of the first element in the destination image - * @param[in] mapx_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapx_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapx_step_x mapx_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapx_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapx_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapx_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] mapy_ptr Pointer to the x remapping image. Supported data types: F32. - * @param[in] mapy_stride_x Stride of the remapping image in X dimension (in bytes) - * @param[in] mapy_step_x mapy_stride_x * number of elements along X processed per work item (in bytes) - * @param[in] mapy_stride_y Stride of the remapping image in Y dimension (in bytes) - * @param[in] mapy_step_y mapy_stride_y * number of elements along Y processed per work item (in bytes) - * @param[in] mapy_offset_first_element_in_bytes Offset of the first element in the remapping image - * @param[in] width Width of the input image - * @param[in] height Height of the input image - * @param[in] border_val Value to use for border around input tensor when in CONSTANT border is selected - */ -__kernel void remap_bilinear_nhwc( - TENSOR4D_DECLARATION(in), - TENSOR4D_DECLARATION(out), - TENSOR4D_DECLARATION(mapx), - TENSOR4D_DECLARATION(mapy), - const float width, - const float height -#ifdef CONSTANT_BORDER - , - const DATA_TYPE border_val -#endif // CONSTANT_BORDER -) -{ - Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(in, 0); - Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT(out, DEPTH_OUT); - Tensor4D mapx = CONVERT_TO_TENSOR4D_STRUCT(mapx, DEPTH_OUT); - Tensor4D mapy = CONVERT_TO_TENSOR4D_STRUCT(mapy, DEPTH_OUT); - - float mapx_coord = (float) * (__global float *)mapx.ptr; - float mapy_coord = (float) * (__global float *)mapy.ptr; - -#ifdef CONSTANT_BORDER - if(mapx_coord < 0 || mapx_coord > width - 1 || mapy_coord < 0 || mapy_coord > height - 1) - { - *((__global DATA_TYPE *)out.ptr) = border_val; - return; - } -#endif // CONSTANT_BORDER - - const float new_xf = floor(mapx_coord); - const float new_yf = floor(mapy_coord); - const float clamped_x = clamp(new_xf, 0.0f, width - 1); - const float clamped_x1 = clamp(new_xf + 1, 0.0f, width - 1); - const float clamped_y = clamp(new_yf, 0.0f, height - 1); - const float clamped_y1 = clamp(new_yf + 1, 0.0f, height - 1); - - float4 ins = (float4)(*((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))), - *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y), (get_global_id(2) / DEPTH_OUT))), - *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x), convert_int(clamped_y1), (get_global_id(2) / DEPTH_OUT))), - *((__global DATA_TYPE *)tensor4D_offset(&in, get_global_id(0), convert_int(clamped_x1), convert_int(clamped_y1), (get_global_id(2) / DEPTH_OUT)))); - - const float a = mapx_coord - new_xf; - const float b = 1.f - a; - const float a1 = mapy_coord - new_yf; - const float b1 = 1.f - a1; - const float fr = ((ins.s0 * b * b1) + (ins.s1 * a * b1) + (ins.s2 * b * a1) + (ins.s3 * a * a1)); - - *((__global DATA_TYPE *)out.ptr) = CONVERT(fr, DATA_TYPE); -} - -#endif // DEPTH_OUT
\ No newline at end of file diff --git a/src/core/CL/kernels/CLRemapKernel.cpp b/src/core/CL/kernels/CLRemapKernel.cpp deleted file mode 100644 index ea3b637e8d..0000000000 --- a/src/core/CL/kernels/CLRemapKernel.cpp +++ /dev/null @@ -1,183 +0,0 @@ -/* - * Copyright (c) 2017-2021 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 "src/core/CL/kernels/CLRemapKernel.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/Helpers.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Validate.h" -#include "src/core/AccessWindowStatic.h" -#include "src/core/helpers/WindowHelpers.h" - -namespace arm_compute -{ -CLRemapKernel::CLRemapKernel() - : _input(nullptr), _output(nullptr), _map_x(nullptr), _map_y(nullptr), _data_layout(DataLayout::NCHW) -{ - _type = CLKernelType::ELEMENTWISE; -} - -BorderSize CLRemapKernel::border_size() const -{ - return _data_layout == DataLayout::NCHW ? BorderSize(1) : BorderSize(0); -} - -template <class T> -void CLRemapKernel::set_constant_border(unsigned int idx, const PixelValue &constant_border_value) -{ - T value; - constant_border_value.get(value); - ICLKernel::add_argument<T>(idx, static_cast<T>(value)); -} - -Status CLRemapKernel::validate(const ITensorInfo *input, const ITensorInfo *map_x, const ITensorInfo *map_y, const ITensorInfo *output, RemapInfo info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, map_x, map_y, output); - if(input->data_layout() == DataLayout::NCHW) - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); - } - else - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::F16); - } - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() != output->data_type(), "Input/output have different data types"); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(map_x, 1, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(map_y, 1, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.policy == InterpolationPolicy::AREA, "Area interpolation is not supported!"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.border_mode != BorderMode::CONSTANT && info.border_mode != BorderMode::UNDEFINED, "Border mode not supported"); - return Status{}; -} - -void CLRemapKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *map_x, const ICLTensor *map_y, ICLTensor *output, RemapInfo info) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, map_x, map_y, output); - ARM_COMPUTE_ERROR_THROW_ON(CLRemapKernel::validate(input->info(), map_x->info(), map_y->info(), output->info(), info)); - - _input = input; - _output = output; - _map_x = map_x; - _map_y = map_y; - _data_layout = input->info()->data_layout(); - - const bool is_nhwc = _data_layout == DataLayout::NHWC; - const bool is_constant_border = info.border_mode == BorderMode::CONSTANT; - - // 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(is_nhwc, "-DDEPTH_OUT=" + support::cpp11::to_string(output->info()->dimension(2))); - build_opts.add_option_if(is_constant_border, "-DCONSTANT_BORDER"); - - const std::string interpolation_name = lower_string(string_from_interpolation_policy(info.policy)); - const std::string kernel_name = "remap_" + interpolation_name + "_" + lower_string(string_from_data_layout(_data_layout)); - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - - const unsigned int num_elems_processed_per_iteration = is_nhwc ? 1 : 4; - const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); - const int idx_width = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH); - const int input_height = input->info()->dimension(idx_height); - const int input_width = input->info()->dimension(idx_width); - - // Configure window - Window win = calculate_max_window(*_output->info(), Steps(num_elems_processed_per_iteration)); - - // Update padding in NCHW case - if(_data_layout == DataLayout::NCHW) - { - const int total_right = ceil_to_multiple(input_width, num_elems_processed_per_iteration); - const int access_right = total_right + (((total_right - input_width) == 0) ? border_size().right : 0); - AccessWindowStatic input_access(input->info(), -border_size().left, -border_size().top, access_right, input_height + border_size().bottom); - - AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); - - update_window_and_padding(win, input_access, output_access); - } - - ICLKernel::configure_internal(win); - - // Set static arguments - unsigned int idx = 4 * (is_nhwc ? num_arguments_per_4D_tensor() : num_arguments_per_2D_tensor()); - _kernel.setArg<cl_float>(idx++, input_width); - _kernel.setArg<cl_float>(idx++, input_height); - if(is_nhwc && is_constant_border) - { - switch(input->info()->data_type()) - { - case DataType::U8: - set_constant_border<uint8_t>(idx, info.constant_border_value); - break; - case DataType::F16: - static_assert(sizeof(cl_half) == sizeof(half), "Half must be same size as cl_half"); - static_assert(sizeof(cl_half) == 2, "Half must be 16 bit"); - set_constant_border<half>(idx, info.constant_border_value); - break; - default: - ARM_COMPUTE_ERROR("Data Type not handled"); - } - } -} - -void CLRemapKernel::run(const Window &window, cl::CommandQueue &queue) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); - switch(_data_layout) - { - case DataLayout::NCHW: - { - Window slice = window.first_slice_window_2D(); - do - { - unsigned int idx = 0; - add_2D_tensor_argument(idx, _input, slice); - add_2D_tensor_argument(idx, _output, slice); - add_2D_tensor_argument(idx, _map_x, slice); - add_2D_tensor_argument(idx, _map_y, slice); - enqueue(queue, *this, slice, lws_hint()); - - } - while(window.slide_window_slice_2D(slice)); - break; - } - case DataLayout::NHWC: - { - Window collapsed = window.collapse(ICLKernel::window(), Window::DimZ); - Window slice = collapsed.first_slice_window_4D(); - - unsigned int idx = 0; - add_4D_tensor_argument(idx, _input, slice); - add_4D_tensor_argument(idx, _output, slice); - add_4D_tensor_argument(idx, _map_x, slice); - add_4D_tensor_argument(idx, _map_y, slice); - enqueue(queue, *this, slice, lws_hint()); - break; - } - default: - ARM_COMPUTE_ERROR("Invalid Data layout"); - } -} -} // namespace arm_compute diff --git a/src/core/CL/kernels/CLRemapKernel.h b/src/core/CL/kernels/CLRemapKernel.h deleted file mode 100644 index 93b0b4e660..0000000000 --- a/src/core/CL/kernels/CLRemapKernel.h +++ /dev/null @@ -1,88 +0,0 @@ -/* - * Copyright (c) 2017-2021 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. - */ -#ifndef ARM_COMPUTE_CLREMAPKERNEL_H -#define ARM_COMPUTE_CLREMAPKERNEL_H - -#include "arm_compute/core/KernelDescriptors.h" -#include "arm_compute/core/Types.h" -#include "src/core/CL/ICLKernel.h" - -namespace arm_compute -{ -class ICLTensor; - -/** OpenCL kernel to perform a remap on a tensor */ -class CLRemapKernel : public ICLKernel -{ -public: - /** Default constructor */ - CLRemapKernel(); - /** Prevent instances of this class from being copied (As this class contains pointers) */ - CLRemapKernel(const CLRemapKernel &) = delete; - /** Prevent instances of this class from being copied (As this class contains pointers) */ - CLRemapKernel &operator=(const CLRemapKernel &) = delete; - /** Allow instances of this class to be moved */ - CLRemapKernel(CLRemapKernel &&) = default; - /** Allow instances of this class to be moved */ - CLRemapKernel &operator=(CLRemapKernel &&) = default; - /** Initialize the kernel's input, output and border mode. - * - * @param[in] compile_context The compile context to be used. - * @param[in] input Source tensor. Data types supported: U8 (or F16 when layout is NHWC). - * @param[in] map_x Map for X coordinates. Data types supported: F32. - * @param[in] map_y Map for Y coordinates. Data types supported: F32. - * @param[out] output Destination tensor. Data types supported: Same as @p input. All but the lowest two dimensions must be the same size as in the input tensor, i.e. remapping is only performed within the XY-plane. - * @param[in] info RemapInfo struct: - * - policy Interpolation policy to use. Only NEAREST and BILINEAR are supported. - * - border_mode Border mode to use on the input tensor. Only CONSTANT and UNDEFINED are supported. - * - constant_border_value Constant value to use for borders if border_mode is set to CONSTANT. - */ - void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *map_x, const ICLTensor *map_y, ICLTensor *output, RemapInfo info); - /** Checks if the kernel's input, output and border mode will lead to a valid configuration of @ref CLRemapKernel - * - * Similar to @ref CLRemapKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *map_x, const ICLTensor *map_y, ICLTensor *output, RemapInfo info) - * - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *map_x, const ITensorInfo *map_y, const ITensorInfo *output, RemapInfo info); - /** Function to set the constant value on fill border kernel depending on type. - * - * @param[in] idx Index of the kernel argument to set. - * @param[in] constant_border_value Constant value to use for borders if border_mode is set to CONSTANT. - */ - template <class T> - void set_constant_border(unsigned int idx, const PixelValue &constant_border_value); - - // Inherited methods overridden: - void run(const Window &window, cl::CommandQueue &queue) override; - BorderSize border_size() const override; - -private: - const ICLTensor *_input; - ICLTensor *_output; - const ICLTensor *_map_x; - const ICLTensor *_map_y; - DataLayout _data_layout; -}; -} // namespace arm_compute -#endif /*ARM_COMPUTE_CLREMAPKERNEL_H */ diff --git a/src/core/NEON/NEKernels.h b/src/core/NEON/NEKernels.h index af301c8d16..cd01659c0d 100644 --- a/src/core/NEON/NEKernels.h +++ b/src/core/NEON/NEKernels.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2021 Arm Limited. + * Copyright (c) 2016-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -55,7 +55,6 @@ #include "src/core/NEON/kernels/NEROIPoolingLayerKernel.h" #include "src/core/NEON/kernels/NERangeKernel.h" #include "src/core/NEON/kernels/NEReductionOperationKernel.h" -#include "src/core/NEON/kernels/NERemapKernel.h" #include "src/core/NEON/kernels/NEReorgLayerKernel.h" #include "src/core/NEON/kernels/NEReverseKernel.h" #include "src/core/NEON/kernels/NESelectKernel.h" diff --git a/src/core/NEON/kernels/NERemapKernel.cpp b/src/core/NEON/kernels/NERemapKernel.cpp deleted file mode 100644 index a1ba29e4c4..0000000000 --- a/src/core/NEON/kernels/NERemapKernel.cpp +++ /dev/null @@ -1,326 +0,0 @@ -/* - * Copyright (c) 2017-2021 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 "src/core/NEON/kernels/NERemapKernel.h" - -#include "arm_compute/core/Error.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/ITensor.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/Window.h" -#include "src/core/AccessWindowStatic.h" -#include "src/core/helpers/AutoConfiguration.h" -#include "src/core/helpers/ScaleHelpers.h" -#include "src/core/helpers/WindowHelpers.h" - -#include <arm_neon.h> -#include <cstddef> -#include <cstdint> - -using namespace arm_compute::scale_helpers; - -namespace arm_compute -{ -class Coordinates; - -namespace -{ -inline int32_t num_out_of_tensor(const float *mapx_ptr, const float *mapy_ptr, const int32x4_t &width_1, const int32x4_t &height_1) -{ - const int32x4_t mapx_s32 = vcvtq_s32_f32(vld1q_f32(mapx_ptr)); - const int32x4_t mapy_s32 = vcvtq_s32_f32(vld1q_f32(mapy_ptr)); - - const int32x4_t outbx_s32 = vminq_s32(vmaxq_s32(vminq_s32(vsubq_s32(width_1, mapx_s32), mapx_s32), vdupq_n_s32(-1)), vdupq_n_s32(0)); // Contains -1 if out of border in x, 0 otherwise - const int32x4_t outby_s32 = vminq_s32(vmaxq_s32(vminq_s32(vsubq_s32(height_1, mapy_s32), mapy_s32), vdupq_n_s32(-1)), vdupq_n_s32(0)); // Contains -1 if out of border in y, 0 otherwise - - const int32x4_t out_of_tensor_v = vminq_s32(outbx_s32, outby_s32); -#if defined(__aarch64__) - // only AArch64 supports vaddv - return vaddvq_s32(out_of_tensor_v); -#else // __aarch64__ - return vgetq_lane_s32(out_of_tensor_v, 0) + vgetq_lane_s32(out_of_tensor_v, 1) + vgetq_lane_s32(out_of_tensor_v, 2) + vgetq_lane_s32(out_of_tensor_v, 3); -#endif // __aarch64__ -} - -inline void serial_remap_nearest_interpolation(const uint8_t *in_ptr, const float *mapx_ptr, const float *mapy_ptr, uint8_t *out_ptr, - int32_t width_val, int32_t height_val, int32_t in_stride_val, uint8_t constant_border_value) -{ - const auto x_s32 = static_cast<int32_t>(*mapx_ptr); - const auto y_s32 = static_cast<int32_t>(*mapy_ptr); - if(x_s32 < 0 || y_s32 < 0 || x_s32 >= width_val || y_s32 >= height_val) - { - *(out_ptr) = constant_border_value; - } - else - { - *(out_ptr) = in_ptr[x_s32 + y_s32 * in_stride_val]; - } -} - -inline int32x4_t offset_nearest_interpolation(const float *mapx_ptr, const float *mapy_ptr, const int32x4_t &stride) -{ - const int32x4_t mapx_s32 = vcvtq_s32_f32(vld1q_f32(mapx_ptr)); - const int32x4_t mapy_s32 = vcvtq_s32_f32(vld1q_f32(mapy_ptr)); - return vmlaq_s32(mapx_s32, mapy_s32, stride); -} - -inline uint8_t pixel_bilinear_c1_clamp(const uint8_t *pixel_ptr, int32_t stride, int32_t width, int32_t height, float x, float y, uint8_t constant_border_value) -{ - x = std::max(-1.f, std::min(x, static_cast<float>(width))); - y = std::max(-1.f, std::min(y, static_cast<float>(height))); - - const int32_t xi = static_cast<int32_t>(std::floor(x)); - const int32_t yi = static_cast<int32_t>(std::floor(y)); - - const float dx = x - static_cast<float>(xi); - const float dy = y - static_cast<float>(yi); - - // Calculating the address won't trigger a segfault in case the value is outside the tensor - // The ternary operator resolves the values in both conditions - const uint8_t *a00 = (xi < 0 || xi >= width || yi < 0 || yi >= height) ? &constant_border_value : (pixel_ptr + xi + yi * stride); - const uint8_t *a01 = (xi + 1 >= width || yi < 0 || yi >= height) ? &constant_border_value : (pixel_ptr + xi + 1 + yi * stride); - const uint8_t *a10 = (xi < 0 || xi >= width || yi + 1 >= height) ? &constant_border_value : (pixel_ptr + xi + yi * stride + stride); - const uint8_t *a11 = (xi + 1 >= width || yi + 1 >= height) ? &constant_border_value : (pixel_ptr + xi + 1 + yi * stride + stride); - - const float dx1 = 1.0f - dx; - const float dy1 = 1.0f - dy; - const float w1 = dx1 * dy1; - const float w2 = dx * dy1; - const float w3 = dx1 * dy; - const float w4 = dx * dy; - - return static_cast<uint8_t>((*a00) * w1 + (*a01) * w2 + (*a10) * w3 + (*a11) * w4); -} -} // namespace - -NERemapKernel::NERemapKernel() - : _func(nullptr), _input(nullptr), _output(nullptr), _map_x(nullptr), _map_y(nullptr), _border_mode(BorderMode::UNDEFINED), _constant_border_value(0) -{ -} - -void NERemapKernel::configure(const ITensor *input, const ITensor *map_x, const ITensor *map_y, ITensor *output, InterpolationPolicy policy, BorderMode border_mode, uint8_t constant_border_value) -{ - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(map_x, 1, DataType::F32); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(map_y, 1, DataType::F32); - - _input = input; - _output = output; - _map_x = map_x; - _map_y = map_y; - _border_mode = border_mode; - _constant_border_value = constant_border_value; - - switch(policy) - { - case InterpolationPolicy::NEAREST_NEIGHBOR: - { - _func = &NERemapKernel::remap_nearest; - break; - } - case InterpolationPolicy::BILINEAR: - { - _func = &NERemapKernel::remap_bilinear; - break; - } - default: - ARM_COMPUTE_ERROR("Unsupported interpolation mode"); - break; - } - - // Configure kernel window - Window win = calculate_max_window(*output->info(), Steps()); - INEKernel::configure(win); -} - -void NERemapKernel::remap_nearest(const Window &window) -{ - // Don't increment in X and Y direction for the input tensor - // A pointer to the start of this plane is needed as base for the precomputed offsets - Window win_in(window); - win_in.set(Window::DimX, Window::Dimension(0, 0, 0)); - win_in.set(Window::DimY, Window::Dimension(0, 0, 0)); - - const auto window_start_x = static_cast<int32_t>(window.x().start()); - const auto window_end_x = static_cast<int32_t>(window.x().end()); - const int32_t window_step_x = 8; - - // Don't increment in X direction for the output, mapx, mapy tensors - Window win(window); - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator in(_input, win_in); - Iterator out(_output, win); - Iterator mapx(_map_x, win); - Iterator mapy(_map_y, win); - - const int32_t width_val = static_cast<int32_t>(_input->info()->dimension(0)); - const int32_t height_val = static_cast<int32_t>(_input->info()->dimension(1)); - const int32_t in_stride_val = static_cast<int32_t>(_input->info()->strides_in_bytes()[1]); - const int32x4_t width_1 = vdupq_n_s32(width_val - 1); - const int32x4_t height_1 = vdupq_n_s32(height_val - 1); - const int32x4_t in_stride = vdupq_n_s32(in_stride_val); - - execute_window_loop(win, [&](const Coordinates &) - { - auto mapx_ptr = reinterpret_cast<const float *>(mapx.ptr()); - auto mapy_ptr = reinterpret_cast<const float *>(mapy.ptr()); - const uint8_t *in_ptr = in.ptr(); - uint8_t *out_ptr = out.ptr(); - int32_t x = window_start_x; - for(; x < window_end_x - window_step_x; x += window_step_x, mapx_ptr += window_step_x, mapy_ptr += window_step_x, out_ptr += window_step_x) - { - const int32_t out_of_tensor0 = num_out_of_tensor(mapx_ptr, mapy_ptr + 0, width_1, height_1); - const int32_t out_of_tensor1 = num_out_of_tensor(mapx_ptr + 4, mapy_ptr + 4, width_1, height_1); - const int32_t out_of_tensor = out_of_tensor0 + out_of_tensor1; - - if(out_of_tensor == -8) - { - // All elements are out of xy plane - uint8x8_t tmp = vdup_n_u8(_constant_border_value); - vst1_u8(out_ptr, tmp); - } - else if(out_of_tensor < 0) - { - // Some elements are out of xy plane - serial_remap_nearest_interpolation(in_ptr, mapx_ptr, mapy_ptr, out_ptr, width_val, height_val, in_stride_val, _constant_border_value); - serial_remap_nearest_interpolation(in_ptr, mapx_ptr + 1, mapy_ptr + 1, out_ptr + 1, width_val, height_val, in_stride_val, _constant_border_value); - serial_remap_nearest_interpolation(in_ptr, mapx_ptr + 2, mapy_ptr + 2, out_ptr + 2, width_val, height_val, in_stride_val, _constant_border_value); - serial_remap_nearest_interpolation(in_ptr, mapx_ptr + 3, mapy_ptr + 3, out_ptr + 3, width_val, height_val, in_stride_val, _constant_border_value); - serial_remap_nearest_interpolation(in_ptr, mapx_ptr + 4, mapy_ptr + 4, out_ptr + 4, width_val, height_val, in_stride_val, _constant_border_value); - serial_remap_nearest_interpolation(in_ptr, mapx_ptr + 5, mapy_ptr + 5, out_ptr + 5, width_val, height_val, in_stride_val, _constant_border_value); - serial_remap_nearest_interpolation(in_ptr, mapx_ptr + 6, mapy_ptr + 6, out_ptr + 6, width_val, height_val, in_stride_val, _constant_border_value); - serial_remap_nearest_interpolation(in_ptr, mapx_ptr + 7, mapy_ptr + 7, out_ptr + 7, width_val, height_val, in_stride_val, _constant_border_value); - } - else - { - // All elements are in xy plane - uint8x8_t tmp = vdup_n_u8(0); - const int32x4_t offset0 = offset_nearest_interpolation(mapx_ptr, mapy_ptr, in_stride); - const int32x4_t offset1 = offset_nearest_interpolation(mapx_ptr + 4, mapy_ptr + 4, in_stride); - tmp = vset_lane_u8(in_ptr[vgetq_lane_s32(offset0, 0)], tmp, 0); - tmp = vset_lane_u8(in_ptr[vgetq_lane_s32(offset0, 1)], tmp, 1); - tmp = vset_lane_u8(in_ptr[vgetq_lane_s32(offset0, 2)], tmp, 2); - tmp = vset_lane_u8(in_ptr[vgetq_lane_s32(offset0, 3)], tmp, 3); - tmp = vset_lane_u8(in_ptr[vgetq_lane_s32(offset1, 0)], tmp, 4); - tmp = vset_lane_u8(in_ptr[vgetq_lane_s32(offset1, 1)], tmp, 5); - tmp = vset_lane_u8(in_ptr[vgetq_lane_s32(offset1, 2)], tmp, 6); - tmp = vset_lane_u8(in_ptr[vgetq_lane_s32(offset1, 3)], tmp, 7); - vst1_u8(out_ptr, tmp); - } - } - for(; x < window_end_x; ++x, ++mapx_ptr, ++mapy_ptr, ++out_ptr) - { - serial_remap_nearest_interpolation(in_ptr, mapx_ptr, mapy_ptr, out_ptr, width_val, height_val, in_stride_val, _constant_border_value); - } - }, - in, out, mapx, mapy); -} - -void NERemapKernel::remap_bilinear(const Window &window) -{ - // Don't increment in X and Y direction for the input tensor - // A pointer to the start of this plane is needed as base for the precomputed offsets - Window win_in(window); - win_in.set(Window::DimX, Window::Dimension(0, 0, 0)); - win_in.set(Window::DimY, Window::Dimension(0, 0, 0)); - - const auto window_start_x = static_cast<int32_t>(window.x().start()); - const auto window_end_x = static_cast<int32_t>(window.x().end()); - const int32_t window_step_x = 8; - - // Don't increment in X direction for the output, mapx, mapy tensors - Window win(window); - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - Iterator in(_input, win_in); - Iterator out(_output, win); - Iterator mapx(_map_x, win); - Iterator mapy(_map_y, win); - - const int32_t width_val = static_cast<int32_t>(_input->info()->dimension(0)); - const int32_t height_val = static_cast<int32_t>(_input->info()->dimension(1)); - const int32x4_t width_2 = vdupq_n_s32(width_val - 2); - const int32x4_t height_2 = vdupq_n_s32(height_val - 2); - const int32_t in_stride_val = static_cast<int32_t>(_input->info()->strides_in_bytes()[1]); - - execute_window_loop(win, [&](const Coordinates &) - { - auto mapx_ptr = reinterpret_cast<const float *>(mapx.ptr()); - auto mapy_ptr = reinterpret_cast<const float *>(mapy.ptr()); - const uint8_t *in_ptr = in.ptr(); - uint8_t *out_ptr = out.ptr(); - int32_t x = window_start_x; - for(; x < window_end_x - window_step_x; x += window_step_x, mapx_ptr += window_step_x, mapy_ptr += window_step_x, out_ptr += window_step_x) - { - const int32_t out_of_tensor0 = num_out_of_tensor(mapx_ptr, mapy_ptr + 0, width_2, height_2); - const int32_t out_of_tensor1 = num_out_of_tensor(mapx_ptr + 4, mapy_ptr + 4, width_2, height_2); - const int32_t out_of_tensor = out_of_tensor0 + out_of_tensor1; - - if(out_of_tensor < 0) - { - // Elements are out of xy plane - *(out_ptr) = pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[0], mapy_ptr[0], _constant_border_value); - *(out_ptr + 1) = pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[1], mapy_ptr[1], _constant_border_value); - *(out_ptr + 2) = pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[2], mapy_ptr[2], _constant_border_value); - *(out_ptr + 3) = pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[3], mapy_ptr[3], _constant_border_value); - *(out_ptr + 4) = pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[4], mapy_ptr[4], _constant_border_value); - *(out_ptr + 5) = pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[5], mapy_ptr[5], _constant_border_value); - *(out_ptr + 6) = pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[6], mapy_ptr[6], _constant_border_value); - *(out_ptr + 7) = pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[7], mapy_ptr[7], _constant_border_value); - } - else - { - // All elements are in xy plane - uint8x8_t tmp = vdup_n_u8(0); - tmp = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[0], mapy_ptr[0], _constant_border_value), tmp, 0); - tmp = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[1], mapy_ptr[1], _constant_border_value), tmp, 1); - tmp = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[2], mapy_ptr[2], _constant_border_value), tmp, 2); - tmp = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[3], mapy_ptr[3], _constant_border_value), tmp, 3); - tmp = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[4], mapy_ptr[4], _constant_border_value), tmp, 4); - tmp = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[5], mapy_ptr[5], _constant_border_value), tmp, 5); - tmp = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[6], mapy_ptr[6], _constant_border_value), tmp, 6); - tmp = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[7], mapy_ptr[7], _constant_border_value), tmp, 7); - vst1_u8(out_ptr, tmp); - } - } - for(; x < window_end_x; ++x, ++mapx_ptr, ++mapy_ptr, ++out_ptr) - { - *(out_ptr) = pixel_bilinear_c1_clamp(in_ptr, in_stride_val, width_val, height_val, mapx_ptr[0], mapy_ptr[0], _constant_border_value); - } - }, - in, out, mapx, mapy); -} - -void NERemapKernel::run(const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_UNUSED(info); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - ARM_COMPUTE_ERROR_ON(_func == nullptr); - - (this->*_func)(window); -} -} // namespace arm_compute
\ No newline at end of file diff --git a/src/core/NEON/kernels/NERemapKernel.h b/src/core/NEON/kernels/NERemapKernel.h deleted file mode 100644 index 33e929805a..0000000000 --- a/src/core/NEON/kernels/NERemapKernel.h +++ /dev/null @@ -1,86 +0,0 @@ -/* - * Copyright (c) 2016-2021 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. - */ -#ifndef ARM_COMPUTE_NEREMAPKERNEL_H -#define ARM_COMPUTE_NEREMAPKERNEL_H - -#include "arm_compute/core/Types.h" -#include "src/core/NEON/INEKernel.h" - -namespace arm_compute -{ -class ITensor; - -/** Kernel to perform a remap on a tensor */ -class NERemapKernel : public INEKernel -{ -public: - const char *name() const override - { - return "NERemapKernel"; - } - /** Default constructor */ - NERemapKernel(); - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NERemapKernel(const NERemapKernel &) = delete; - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NERemapKernel &operator=(const NERemapKernel &) = delete; - /** Allow instances of this class to be moved */ - NERemapKernel(NERemapKernel &&) = default; - /** Allow instances of this class to be moved */ - NERemapKernel &operator=(NERemapKernel &&) = default; - /** Default destructor */ - ~NERemapKernel() = default; - - /** Initialize the kernel's input, output and border mode. - * - * @param[in] input Source tensor. Data type supported: U8. - * @param[in] map_x Map for X coordinates. Data type supported: F32. - * @param[in] map_y Map for Y coordinates. Data type supported: F32. - * @param[out] output Destination tensor. Data types supported: U8. All but the lowest two dimensions must be the same size as in the input tensor, i.e. remapping is only performed within the XY-plane. - * @param[in] policy The interpolation type. - * @param[in] border_mode Border mode to use on the input tensor. - * @param[in] constant_border_value (Optional) Constant value to use for borders if border_mode is set to CONSTANT. Defaults to 0. - */ - void configure(const ITensor *input, const ITensor *map_x, const ITensor *map_y, ITensor *output, InterpolationPolicy policy, BorderMode border_mode, uint8_t constant_border_value = 0); - - // Inherited methods overridden: - void run(const Window &window, const ThreadInfo &info) override; - -private: - /** function to perform nearest interpolation on the given window */ - void remap_nearest(const Window &window); - /** function to perform bilinear interpolation on the given window */ - void remap_bilinear(const Window &window); - /** Remap function to use for the particular interpolation type passed to configure() */ - void (NERemapKernel::*_func)(const Window &window); - - const ITensor *_input; /**< Input image */ - ITensor *_output; /**< Output image */ - const ITensor *_map_x; /**< Input remap x coordinates */ - const ITensor *_map_y; /**< Input remap y coordinates */ - BorderMode _border_mode; /**< Border mode */ - uint8_t _constant_border_value; /**< Border value to use */ -}; -} // namespace arm_compute -#endif /*ARM_COMPUTE_NEREMAPKERNEL_H */
\ No newline at end of file |