Port Arm(R) Neon(TM) Scale to new API
Partially resolves: COMPMID-4190
Change-Id: I0c1e32ff6176775c9b7bf547899a791fd318ba0a
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5192
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: TeresaARM <teresa.charlinreyes@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
diff --git a/src/core/cpu/kernels/CpuScaleKernel.cpp b/src/core/cpu/kernels/CpuScaleKernel.cpp
new file mode 100644
index 0000000..22d1332
--- /dev/null
+++ b/src/core/cpu/kernels/CpuScaleKernel.cpp
@@ -0,0 +1,618 @@
+/*
+ * 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.
+ */
+#include "src/core/cpu/kernels/CpuScaleKernel.h"
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/Utility.h"
+#include "src/core/AccessWindowStatic.h"
+#include "src/core/CPP/Validate.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Registrars.h"
+#include "src/core/cpu/kernels/scale/neon/list.h"
+#include "src/core/cpu/kernels/scale/sve/list.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/helpers/WindowHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <arm_neon.h>
+#include <map>
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+namespace
+{
+struct ScaleSelectorData
+{
+ DataType dt;
+};
+using ScaleSelectorPtr = std::add_pointer<bool(const ScaleSelectorData &data)>::type;
+using ScaleKernelPtr = std::add_pointer<void(const ITensor *, ITensor *, const ITensor *, const ITensor *, const ITensor *,
+ InterpolationPolicy, BorderMode, PixelValue, float, bool, const Window &)>::type;
+struct ScaleKernel
+{
+ const char *name;
+ const ScaleSelectorPtr is_selected;
+ ScaleKernelPtr ukernel;
+};
+
+static const ScaleKernel available_kernels[] =
+{
+#if defined(__ARM_FEATURE_SVE)
+ {
+ "fp16_sve_scale",
+ [](const ScaleSelectorData & data) { return data.dt == DataType::F16; },
+ REGISTER_FP16_NEON(arm_compute::cpu::fp16_sve_scale)
+ },
+ {
+ "f32_sve_scale",
+ [](const ScaleSelectorData & data) { return data.dt == DataType::F32; },
+ REGISTER_FP32_NEON(arm_compute::cpu::fp32_sve_scale)
+ },
+ {
+ "qasymm8_sve_scale",
+ [](const ScaleSelectorData & data) { return data.dt == DataType::QASYMM8; },
+ REGISTER_QASYMM8_NEON(arm_compute::cpu::qasymm8_sve_scale)
+ },
+ {
+ "qasymm8_signed_sve_scale",
+ [](const ScaleSelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; },
+ REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::qasymm8_signed_sve_scale)
+ },
+ {
+ "u8_sve_scale",
+ [](const ScaleSelectorData & data) { return data.dt == DataType::U8; },
+ REGISTER_INTEGER_NEON(arm_compute::cpu::u8_sve_scale)
+ },
+ {
+ "s16_sve_scale",
+ [](const ScaleSelectorData & data) { return data.dt == DataType::S16; },
+ REGISTER_INTEGER_NEON(arm_compute::cpu::s16_sve_scale)
+ },
+#else /* !defined(__ARM_FEATURE_SVE) */
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+ {
+ "common_neon_scale",
+ [](const ScaleSelectorData & data) { return data.dt == DataType::F16; },
+ REGISTER_FP16_NEON(arm_compute::cpu::common_neon_scale<float16_t>)
+ },
+#endif /* !defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) */
+ {
+ "common_neon_scale",
+ [](const ScaleSelectorData & data) { return data.dt == DataType::F32; },
+ REGISTER_FP32_NEON(arm_compute::cpu::common_neon_scale<float>)
+ },
+ {
+ "qasymm8_neon_scale",
+ [](const ScaleSelectorData & data) { return data.dt == DataType::QASYMM8; },
+ REGISTER_QASYMM8_NEON(arm_compute::cpu::qasymm8_neon_scale)
+ },
+ {
+ "qasymm8_signed_neon_scale",
+ [](const ScaleSelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; },
+ REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::qasymm8_signed_neon_scale)
+ },
+ {
+ "common_neon_scale",
+ [](const ScaleSelectorData & data) { return data.dt == DataType::U8; },
+ REGISTER_INTEGER_NEON(arm_compute::cpu::common_neon_scale<uint8_t>)
+ },
+ {
+ "common_neon_scale",
+ [](const ScaleSelectorData & data) { return data.dt == DataType::S16; },
+ REGISTER_INTEGER_NEON(arm_compute::cpu::common_neon_scale<int16_t>)
+ },
+#endif /* !defined(__ARM_FEATURE_SVE) */
+};
+
+/** Micro-kernel selector
+ *
+ * @param[in] data Selection data passed to help pick the appropriate micro-kernel
+ *
+ * @return A matching micro-kernel else nullptr
+ */
+const ScaleKernel *get_implementation(const ScaleSelectorData &data)
+{
+ for(const auto &uk : available_kernels)
+ {
+ if(uk.is_selected(data))
+ {
+ return &uk;
+ }
+ }
+ return nullptr;
+}
+
+Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dx, const ITensorInfo *dy,
+ const ITensorInfo *offsets, ITensorInfo *dst, const ScaleKernelInfo &info)
+{
+ const auto *uk = get_implementation(ScaleSelectorData{ src->data_type() });
+ ARM_COMPUTE_RETURN_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
+
+ ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(dst);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
+ ARM_COMPUTE_RETURN_ERROR_ON(dst == src);
+ ARM_COMPUTE_RETURN_ERROR_ON(info.sampling_policy != SamplingPolicy::CENTER && info.sampling_policy != SamplingPolicy::TOP_LEFT);
+ ARM_COMPUTE_UNUSED(info.constant_border_value);
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.use_padding, "Padding is not supported");
+
+ const DataLayout data_layout = info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : info.data_layout;
+ const auto width_index = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+ const auto height_index = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+ const auto output_width = dst->dimension(width_index);
+ const auto output_height = dst->dimension(height_index);
+ ARM_COMPUTE_RETURN_ERROR_ON(output_width == 0);
+ ARM_COMPUTE_RETURN_ERROR_ON(output_height == 0);
+
+ if(info.interpolation_policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(offsets, 1, DataType::S32);
+ }
+
+ if(info.interpolation_policy == InterpolationPolicy::BILINEAR)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(offsets, 1, DataType::S32);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dx, 1, DataType::F32);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dy, 1, DataType::F32);
+ }
+
+ ARM_COMPUTE_RETURN_ERROR_ON(info.align_corners && !scale_utils::is_align_corners_allowed_sampling_policy(info.sampling_policy));
+
+ if(info.interpolation_policy == InterpolationPolicy::AREA)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON(data_layout != DataLayout::NCHW);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::U8);
+ }
+
+ return Status{};
+}
+} // namespace
+
+CpuScaleKernel::CpuScaleKernel()
+ : _func(nullptr), _policy(), _border_mode(), _constant_border_value(PixelValue()), _sampling_offset(0), _align_corners(false), _data_layout(DataLayout::UNKNOWN)
+{
+}
+
+void CpuScaleKernel::configure(const ITensorInfo *src, const ITensorInfo *dx, const ITensorInfo *dy, const ITensorInfo *offsets,
+ ITensorInfo *dst, const ScaleKernelInfo &info)
+{
+ ARM_COMPUTE_UNUSED(dx, dy, offsets);
+ ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+ // Perform validation step
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src,
+ dx,
+ dy,
+ offsets,
+ dst,
+ info));
+
+ // Get data layout and width/height indices
+ _data_layout = info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : info.data_layout;
+ const int idx_width = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
+ const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+
+ _policy = info.interpolation_policy;
+ _border_mode = info.border_mode;
+ _constant_border_value = info.constant_border_value;
+ _align_corners = info.align_corners;
+
+ if(info.sampling_policy == SamplingPolicy::CENTER)
+ {
+ _sampling_offset = 0.5f;
+ }
+
+ // Compute the ratio between source width/height and destination width/height
+ const auto wr = scale_utils::calculate_resize_ratio(src->dimension(idx_width), dst->dimension(idx_width), _align_corners);
+ const auto hr = scale_utils::calculate_resize_ratio(src->dimension(idx_height), dst->dimension(idx_height), _align_corners);
+
+ // Area interpolation behaves as Nearest Neighbour in case of up-sampling
+ _policy = (_policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f) ? InterpolationPolicy::NEAREST_NEIGHBOR : _policy;
+
+ if(_border_mode == BorderMode::UNDEFINED)
+ {
+ _border_mode = BorderMode::CONSTANT;
+ _constant_border_value = PixelValue();
+ }
+
+#ifdef ENABLE_NCHW_KERNELS
+ // Configure scale function to run
+ if(_data_layout == DataLayout::NCHW)
+ {
+ std::string function_to_call("scale_");
+ function_to_call += string_from_data_type(src->data_type()) + "_";
+ function_to_call += string_from_data_layout(_data_layout) + "_";
+ function_to_call += string_from_interpolation_policy(_policy);
+
+ static std::map<std::string, ScaleFunctionPtr> map_function =
+ {
+ { "scale_U8_NCHW_AREA_CONSTANT", &CpuScaleKernel::scale_area_nchw_u8 },
+
+ { "scale_U8_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_nchw<uint8_t> },
+ { "scale_U8_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<uint8_t> },
+
+ { "scale_QASYMM8_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_qasymm<uint8_t> },
+ { "scale_QASYMM8_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<uint8_t> },
+
+ { "scale_QASYMM8_SIGNED_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_qasymm<int8_t> },
+ { "scale_QASYMM8_SIGNED_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<int8_t> },
+
+ { "scale_S16_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_nchw<int16_t> },
+ { "scale_S16_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<int16_t> },
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+ { "scale_F16_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_nchw<float16_t> },
+ { "scale_F16_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<float16_t> },
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+
+ { "scale_F32_NCHW_BILINEAR", &CpuScaleKernel::scale_bilinear_nchw<float> },
+ { "scale_F32_NCHW_NEAREST_NEIGHBOUR", &CpuScaleKernel::scale_nearest_nchw<float> },
+ };
+ auto it = map_function.find(function_to_call);
+ if(it != map_function.end())
+ {
+ _func = it->second;
+ }
+ }
+#endif // ENABLE_NCHW_KERNELS
+
+ // Configure window
+ Window win = calculate_max_window(*dst, Steps());
+ ICpuKernel::configure(win);
+}
+
+#ifdef ENABLE_NCHW_KERNELS
+template <typename T>
+void CpuScaleKernel::scale_nearest_nchw(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window)
+{
+ ARM_COMPUTE_UNUSED(dx, dy);
+ const size_t in_stride_x = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(1), dst->info()->dimension(1), _align_corners);
+
+ // 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));
+
+ // Set offsets window
+ Window win_off;
+ win_off.set(Window::DimX, window[Window::DimX]);
+ win_off.set(Window::DimY, window[Window::DimY]);
+ for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
+ {
+ win_off.set(d, Window::Dimension(0, 0, 0));
+ }
+
+ // Create iterators
+ Iterator src_i(src, win_in);
+ Iterator dst_i(dst, window);
+ Iterator offsets_i(offsets, win_off);
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offsets_ptr = reinterpret_cast<const int32_t *>(offsets_i.ptr());
+ const auto in_yi = static_cast<int32_t>(_align_corners ? utils::rounding::round_half_away_from_zero((id.y() + _sampling_offset) * hr) : std::floor((
+ id.y() + _sampling_offset)
+ * hr));
+ const int32_t offset_row = in_yi * in_stride_x;
+ *reinterpret_cast<T *>(dst_i.ptr()) = *(reinterpret_cast<const T *>(src_i.ptr()) + offsets_ptr[0] + offset_row);
+ },
+ src_i, offsets_i, dst_i);
+}
+
+template <typename T>
+void CpuScaleKernel::scale_bilinear_nchw(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window)
+{
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(1), dst->info()->dimension(1), _align_corners);
+ Window win_off;
+ win_off.set(Window::DimX, window.x());
+ win_off.set(Window::DimY, window.y());
+
+ // 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));
+
+ for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
+ {
+ win_off.set(d, Window::Dimension(0, 0, 0));
+ }
+
+ Iterator src_i(src, win_in);
+ Iterator dst_i(dst, window);
+ Iterator offsets_i(offsets, win_off);
+ Iterator dx_i(dx, win_off);
+ Iterator dy_i(dy, win_off);
+
+ const int32_t in_dim_w = src->info()->dimension(0);
+ const int32_t in_dim_h = src->info()->dimension(1);
+ const int32_t in_stride_w = in_dim_w + src->info()->padding().left + src->info()->padding().right;
+
+ if(_border_mode == BorderMode::CONSTANT)
+ {
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+ using ConstType = typename std::conditional<std::is_same<T, float16_t>::value, half, T>::type;
+#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+ using ConstType = T;
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+ const T const_border_value = static_cast<T>(_constant_border_value.get<ConstType>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int32_t index_h = std::floor((id.y() + _sampling_offset) * hr - _sampling_offset);
+ const auto index_w = *(reinterpret_cast<const int32_t *>(offsets_i.ptr()));
+ const auto dx_val = *(reinterpret_cast<const float *>(dx_i.ptr()));
+ const auto dy_val = *(reinterpret_cast<const float *>(dy_i.ptr()));
+ const auto pixel_row_ptr = reinterpret_cast<const T *>(src_i.ptr());
+
+ const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ? (*(pixel_row_ptr + index_w + index_h * in_stride_w)) : const_border_value;
+ const auto a01 = (-1 <= index_w && index_w < in_dim_w - 1 && 0 <= index_h && index_h < in_dim_h) ? (*(pixel_row_ptr + index_w + 1 + index_h * in_stride_w)) : const_border_value;
+ const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h
+ && index_h < in_dim_h - 1) ?
+ (*(pixel_row_ptr + index_w + index_h * in_stride_w + in_stride_w)) :
+ const_border_value;
+ const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h
+ && index_h < in_dim_h - 1) ?
+ (*(pixel_row_ptr + index_w + 1 + index_h * in_stride_w + in_stride_w)) :
+ const_border_value;
+
+ *reinterpret_cast<T *>(dst_i.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ src_i, offsets_i, dx_i, dy_i, dst_i);
+ }
+ else if(_border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int index_h = std::floor((id.y() + _sampling_offset) * hr - _sampling_offset);
+ const auto index_w = *(reinterpret_cast<const int32_t *>(offsets_i.ptr()));
+ const auto dx_val = *(reinterpret_cast<const float *>(dx_i.ptr()));
+ const auto dy_val = *(reinterpret_cast<const float *>(dy_i.ptr()));
+ const auto pixel_row_ptr = reinterpret_cast<const T *>(src_i.ptr());
+
+ auto clamped_x = utility::clamp<int>(index_w, 0, in_dim_w - 1);
+ auto clamped_x1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
+ auto clamped_y = utility::clamp<int>(index_h, 0, in_dim_h - 1);
+ auto clamped_y1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(pixel_row_ptr + clamped_x + clamped_y * in_stride_w);
+ const auto a01 = *(pixel_row_ptr + clamped_x1 + clamped_y * in_stride_w);
+ const auto a10 = *(pixel_row_ptr + clamped_x + clamped_y1 * in_stride_w);
+ const auto a11 = *(pixel_row_ptr + clamped_x1 + clamped_y1 * in_stride_w);
+
+ *reinterpret_cast<T *>(dst_i.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ src_i, offsets_i, dx_i, dy_i, dst_i);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+
+void CpuScaleKernel::scale_area_nchw_u8(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window)
+{
+ ARM_COMPUTE_UNUSED(dx, dy, offsets);
+ using namespace scale_helpers;
+
+ ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::U8);
+
+ // Don't increment in width/height/channels 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));
+ win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+
+ Iterator src_i(src, win_in);
+ Iterator dst_i(dst, window);
+
+ const auto wr = scale_utils::calculate_resize_ratio(src->info()->dimension(0), dst->info()->dimension(0), _align_corners);
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(1), dst->info()->dimension(1), _align_corners);
+ const auto w = src->info()->dimension(0);
+ const auto h = src->info()->dimension(1);
+ const size_t in_stride = src->info()->strides_in_bytes()[1];
+
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto in_ptr = reinterpret_cast<const uint8_t *>(src_i.ptr());
+
+ uint8x8_t tmp0 = vdup_n_u8(0);
+ tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x(), id.y()), tmp0, 0);
+ tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 1, id.y()), tmp0, 1);
+ tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 2, id.y()), tmp0, 2);
+ tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 3, id.y()), tmp0, 3);
+ tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 4, id.y()), tmp0, 4);
+ tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 5, id.y()), tmp0, 5);
+ tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 6, id.y()), tmp0, 6);
+ tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 7, id.y()), tmp0, 7);
+
+ uint8x8_t tmp1 = vdup_n_u8(0);
+ tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 8, id.y()), tmp1, 0);
+ tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 9, id.y()), tmp1, 1);
+ tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 10, id.y()), tmp1, 2);
+ tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 11, id.y()), tmp1, 3);
+ tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 12, id.y()), tmp1, 4);
+ tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 13, id.y()), tmp1, 5);
+ tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 14, id.y()), tmp1, 6);
+ tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 15, id.y()), tmp1, 7);
+
+ vst1q_u8(dst_i.ptr(), vcombine_u8(tmp0, tmp1));
+ },
+ src_i, dst_i);
+}
+
+template <typename T>
+void CpuScaleKernel::scale_bilinear_qasymm(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window)
+{
+ // Get data layout and width/height indices
+ const int idx_width = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
+ const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(idx_height), dst->info()->dimension(idx_height), _align_corners);
+ Window win_off;
+ win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
+ win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+ // 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(idx_width, Window::Dimension(0, 0, 0));
+ win_in.set(idx_height, Window::Dimension(0, 0, 0));
+
+ for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
+ {
+ win_off.set(d, Window::Dimension(0, 0, 0));
+ }
+
+ Iterator src_i(src, win_in);
+ Iterator dst_i(dst, window);
+
+ const int32_t in_dim_w = src->info()->dimension(idx_width);
+ const int32_t in_dim_h = src->info()->dimension(idx_height);
+ const int32_t stride_w = src->info()->strides_in_bytes()[idx_width];
+ const int32_t stride_h = src->info()->strides_in_bytes()[idx_height];
+
+ const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
+ const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
+
+ if(_border_mode == BorderMode::CONSTANT)
+ {
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+ using ConstType = typename std::conditional<std::is_same<T, float16_t>::value, half, T>::type;
+#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+ using ConstType = T;
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+ const T const_border_value = static_cast<T>(_constant_border_value.get<ConstType>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int32_t index_h = std::floor((id[idx_height] + _sampling_offset) * hr - _sampling_offset);
+ const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto pixel_row_ptr = reinterpret_cast<const T *>(src_i.ptr());
+
+ const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ?
+ (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
+ const_border_value;
+ const auto a01 = (-1 <= index_w && index_w < in_dim_w - 1 && 0 <= index_h && index_h < in_dim_h) ?
+ (*(pixel_row_ptr + (index_w + 1) * stride_w + index_h * stride_h)) :
+ const_border_value;
+ const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ?
+ (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) :
+ const_border_value;
+ const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ?
+ (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
+ const_border_value;
+
+ const float inp00 = Qasymm8QuantizationHelper<T>::dequantize(a00, iq_info);
+ const float inp01 = Qasymm8QuantizationHelper<T>::dequantize(a01, iq_info);
+ const float inp10 = Qasymm8QuantizationHelper<T>::dequantize(a10, iq_info);
+ const float inp11 = Qasymm8QuantizationHelper<T>::dequantize(a11, iq_info);
+ *reinterpret_cast<T *>(dst_i.ptr()) = Qasymm8QuantizationHelper<T>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+ },
+ src_i, dst_i);
+ }
+ else if(_border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int index_h = std::floor((id[idx_height] + _sampling_offset) * hr - _sampling_offset);
+ const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto pixel_row_ptr = reinterpret_cast<const T *>(src_i.ptr());
+
+ auto clamped_w = utility::clamp<int>(index_w, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(index_h, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h);
+ const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h);
+ const auto a10 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h1 * stride_h);
+ const auto a11 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h1 * stride_h);
+
+ const float inp00 = Qasymm8QuantizationHelper<T>::dequantize(a00, iq_info);
+ const float inp01 = Qasymm8QuantizationHelper<T>::dequantize(a01, iq_info);
+ const float inp10 = Qasymm8QuantizationHelper<T>::dequantize(a10, iq_info);
+ const float inp11 = Qasymm8QuantizationHelper<T>::dequantize(a11, iq_info);
+ *reinterpret_cast<T *>(dst_i.ptr()) = Qasymm8QuantizationHelper<T>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+ },
+ src_i, dst_i);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+#endif // ENABLE_NCHW_KERNELS
+
+Status CpuScaleKernel::validate(const ITensorInfo *input, const ITensorInfo *dx, const ITensorInfo *dy,
+ const ITensorInfo *offsets, ITensorInfo *output, const ScaleKernelInfo &info)
+{
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, dx, dy, offsets, output, info));
+ return Status{};
+}
+
+void CpuScaleKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
+{
+ ARM_COMPUTE_UNUSED(info);
+ ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+ ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window);
+ ARM_COMPUTE_ERROR_ON(_func == nullptr && _data_layout == DataLayout::NCHW);
+
+ const auto src = tensors.get_const_tensor(TensorType::ACL_SRC);
+ auto dst = tensors.get_tensor(TensorType::ACL_DST);
+ const auto dx = tensors.get_const_tensor(TensorType::ACL_INT_0);
+ const auto dy = tensors.get_const_tensor(TensorType::ACL_INT_1);
+ const auto offsets = tensors.get_const_tensor(TensorType::ACL_INT_2);
+
+ if(_data_layout == DataLayout::NCHW)
+ {
+ (this->*_func)(src, dst, dx, dy, offsets, window);
+ }
+ else
+ {
+ const auto *uk = get_implementation(ScaleSelectorData{ src->info()->data_type() });
+ uk->ukernel(src, dst, offsets, dx, dy, _policy, _border_mode, _constant_border_value, _sampling_offset, _align_corners, window);
+ }
+}
+
+const char *CpuScaleKernel::name() const
+{
+ return "CpuScaleKernel";
+}
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/core/cpu/kernels/CpuScaleKernel.h b/src/core/cpu/kernels/CpuScaleKernel.h
new file mode 100644
index 0000000..c1de8e0
--- /dev/null
+++ b/src/core/cpu/kernels/CpuScaleKernel.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 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_CPU_SCALEKERNEL_H
+#define ARM_COMPUTE_CPU_SCALEKERNEL_H
+
+#include "arm_compute/core/KernelDescriptors.h"
+#include "src/core/common/Macros.h"
+#include "src/core/cpu/ICpuKernel.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace kernels
+{
+/** Arm(R) Neon(TM) kernel to perform scaling on a tensor */
+class CpuScaleKernel : public ICpuKernel
+{
+public:
+ /** Default constructor */
+ CpuScaleKernel();
+ ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuScaleKernel);
+ /** Initialise the kernel's inputs, output and interpolation policy
+ *
+ * @note dx, dy and offsets have the same dimensions (width and height) of the output tensor
+ * @note Using @p policy Area only supports data layout NCHW and input data type U8.
+ *
+ * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/U8/S16/F16/F32.
+ * @param[in] dx Distance x tensor info. Pixel's distance between the X real coordinate and the smallest X following integer. Data type supported: F32
+ * @param[in] dy Distance y tensor info. Pixel's distance between the Y real coordinate and the smallest Y following integer. Data type supported: F32
+ * @param[in] offsets Offset tensor info. Offset to access the pixel with NEAREST interpolation or the top-left pixel with BILINEAR interpolation in the input tensor. Data type supported: S32.
+ * @param[out] dst Destination tensor info. 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. scaling is only performed within the XY-plane.
+ * @param[in] info @ref ScaleKernelInfo to use for configuration
+ */
+ void configure(const ITensorInfo *src, const ITensorInfo *dx, const ITensorInfo *dy, const ITensorInfo *offsets, ITensorInfo *dst,
+ const ScaleKernelInfo &info);
+ /** Static function to check if given info will lead to a valid configuration of @ref CpuScaleKernel
+ *
+ * @note dx, dy and offsets have the same dimensions (width and height) of the output tensor
+ * @note Using @p policy Area only supports data layout NCHW and input data type U8.
+ *
+ * @param[in] src Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/U8/S16/F16/F32.
+ * @param[in] dx Distance x tensor info. Pixel's distance between the X real coordinate and the smallest X following integer. Data type supported: F32
+ * @param[in] dy Distance y tensor info. Pixel's distance between the Y real coordinate and the smallest Y following integer. Data type supported: F32
+ * @param[in] offsets Offset tensor info. Offset to access the pixel with NEAREST interpolation or the top-left pixel with BILINEAR interpolation in the input tensor. Data type supported: S32.
+ * @param[in] dst Destination tensor info. 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. scaling is only performed within the XY-plane.
+ * @param[in] info @ref ScaleKernelInfo to use for validation
+ */
+ static Status validate(const ITensorInfo *src, const ITensorInfo *dx, const ITensorInfo *dy, const ITensorInfo *offsets, ITensorInfo *dst,
+ const ScaleKernelInfo &info);
+
+ // Inherited methods overridden:
+ void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
+ const char *name() const override;
+
+private:
+#ifdef ENABLE_NCHW_KERNELS
+ /** function to perform scale using area interpolation on the given window
+ *
+ * @note Used only in case down-sampling.
+ */
+ void scale_area_nchw_u8(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window);
+
+ /** function to perform scale using bilinear interpolation on the given window */
+ template <typename T>
+ void scale_bilinear_nchw(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window);
+ /** function to perform scale using bilinear interpolation on the given window */
+ template <typename T>
+ void scale_bilinear_qasymm(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window);
+
+ /** function to perform scale using nearest neighbour on the given window */
+ template <typename T>
+ void scale_nearest_nchw(const ITensor *src, ITensor *dst, const ITensor *dx, const ITensor *dy, const ITensor *offsets, const Window &window);
+#endif // ENABLE_NCHW_KERNELS
+
+ /** Scale function to use for the particular function to use */
+ using ScaleFunctionPtr = void (CpuScaleKernel::*)(const ITensor *, ITensor *, const ITensor *, const ITensor *, const ITensor *, const Window &window);
+
+ ScaleFunctionPtr _func;
+ InterpolationPolicy _policy;
+ BorderMode _border_mode;
+ PixelValue _constant_border_value;
+ float _sampling_offset;
+ bool _align_corners;
+ DataLayout _data_layout;
+};
+} // namespace kernels
+} // namespace cpu
+} // namespace arm_compute
+#endif /*ARM_COMPUTE_CPU_SCALEKERNEL_H */
diff --git a/src/core/cpu/kernels/scale/neon/fp16.cpp b/src/core/cpu/kernels/scale/neon/fp16.cpp
new file mode 100644
index 0000000..0ad66ca
--- /dev/null
+++ b/src/core/cpu/kernels/scale/neon/fp16.cpp
@@ -0,0 +1,174 @@
+/*
+ * Copyright (c) 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+
+namespace arm_compute
+{
+namespace
+{
+void fp16_neon_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+ float sampling_offset, bool align_corners, const Window &window)
+{
+ const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+ const size_t in_stride_wc = in_stride_w * in_stride_c;
+ const size_t in_dim_h = src->info()->dimension(2);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+ 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 int window_step_x = 8;
+
+ Window win(window);
+ win.set(Window::DimX, Window::Dimension(0, 1, 1));
+ Iterator out(dst, win);
+
+ const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
+ const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+ execute_window_loop(win, [&](const Coordinates & id)
+ {
+ const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+ const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+ const int offset_row = in_hi * in_stride_wc;
+ int32_t x = window_start_x;
+ const float16_t *in_ptr = reinterpret_cast<const float16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+
+ for(; x <= window_end_x - window_step_x; x += window_step_x)
+ {
+ wrapper::vstore(reinterpret_cast<float16_t *>(out.ptr()) + x,
+ wrapper::vloadq(in_ptr + offset + offset_row + x));
+ }
+ for(; x < window_end_x; ++x)
+ {
+ *(reinterpret_cast<float16_t *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x);
+ }
+ },
+ out);
+}
+
+void fp16_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+ Iterator out(dst, window);
+ const int in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const int in_dim_w = src->info()->dimension(1);
+ const int in_dim_h = src->info()->dimension(2);
+ const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+ // Don't increment in Y and Z 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::DimY, Window::Dimension(0, 0, 0));
+ win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+ Iterator in(src, win_in);
+
+ if(border_mode == BorderMode::CONSTANT)
+ {
+ using ConstType = typename std::conditional<std::is_same<float16_t, float16_t>::value, half, float16_t>::type;
+
+ const float16_t const_border_value = static_cast<float16_t>(constant_border_value.get<ConstType>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+ const float16_t *in_ptr = reinterpret_cast<const float16_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+ const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+ const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+ const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+ const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+ *reinterpret_cast<float16_t *>(out.ptr()) = static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else if(border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+ auto clamped_w = utility::clamp<int>(offset, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+ const auto a01 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+ const auto a10 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+ const auto a11 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+ *reinterpret_cast<float16_t *>(out.ptr()) = static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+}
+namespace cpu
+{
+void fp16_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ if(policy == InterpolationPolicy::BILINEAR)
+ {
+ fp16_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+ }
+ else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ fp16_neon_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+ }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
\ No newline at end of file
diff --git a/src/core/cpu/kernels/scale/neon/integer.cpp b/src/core/cpu/kernels/scale/neon/integer.cpp
new file mode 100644
index 0000000..a2359aa
--- /dev/null
+++ b/src/core/cpu/kernels/scale/neon/integer.cpp
@@ -0,0 +1,293 @@
+/*
+ * Copyright (c) 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+namespace arm_compute
+{
+namespace
+{
+void u8_neon_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+ float sampling_offset, bool align_corners, const Window &window)
+{
+ const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+ const size_t in_stride_wc = in_stride_w * in_stride_c;
+ const size_t in_dim_h = src->info()->dimension(2);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+ 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 int window_step_x = 16;
+
+ Window win(window);
+ win.set(Window::DimX, Window::Dimension(0, 1, 1));
+ Iterator out(dst, win);
+
+ const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
+ const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+ execute_window_loop(win, [&](const Coordinates & id)
+ {
+ const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+ const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+ const int offset_row = in_hi * in_stride_wc;
+ int32_t x = window_start_x;
+ const uint8_t *in_ptr = reinterpret_cast<const uint8_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+
+ for(; x <= window_end_x - window_step_x; x += window_step_x)
+ {
+ wrapper::vstore(reinterpret_cast<uint8_t *>(out.ptr()) + x,
+ wrapper::vloadq(in_ptr + offset + offset_row + x));
+ }
+ for(; x < window_end_x; ++x)
+ {
+ *(reinterpret_cast<uint8_t *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x);
+ }
+ },
+ out);
+}
+
+void u8_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+ Iterator out(dst, window);
+ const int in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const int in_dim_w = src->info()->dimension(1);
+ const int in_dim_h = src->info()->dimension(2);
+ const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+ // Don't increment in Y and Z 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::DimY, Window::Dimension(0, 0, 0));
+ win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+ Iterator in(src, win_in);
+
+ if(border_mode == BorderMode::CONSTANT)
+ {
+ const uint8_t const_border_value = static_cast<uint8_t>(constant_border_value.get<uint8_t>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+ const uint8_t *in_ptr = reinterpret_cast<const uint8_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+ const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+ const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+ const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+ const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+ *reinterpret_cast<uint8_t *>(out.ptr()) = static_cast<uint8_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else if(border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+ auto clamped_w = utility::clamp<int>(offset, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+ const auto a01 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+ const auto a10 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+ const auto a11 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+ *reinterpret_cast<uint8_t *>(out.ptr()) = static_cast<uint8_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+
+void s16_neon_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+ float sampling_offset, bool align_corners, const Window &window)
+{
+ const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+ const size_t in_stride_wc = in_stride_w * in_stride_c;
+ const size_t in_dim_h = src->info()->dimension(2);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+ 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 int window_step_x = 8;
+
+ Window win(window);
+ win.set(Window::DimX, Window::Dimension(0, 1, 1));
+ Iterator out(dst, win);
+
+ const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
+ const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+ execute_window_loop(win, [&](const Coordinates & id)
+ {
+ const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+ const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+ const int offset_row = in_hi * in_stride_wc;
+ int32_t x = window_start_x;
+ const int16_t *in_ptr = reinterpret_cast<const int16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+
+ for(; x <= window_end_x - window_step_x; x += window_step_x)
+ {
+ wrapper::vstore(reinterpret_cast<int16_t *>(out.ptr()) + x,
+ wrapper::vloadq(in_ptr + offset + offset_row + x));
+ }
+ for(; x < window_end_x; ++x)
+ {
+ *(reinterpret_cast<int16_t *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x);
+ }
+ },
+ out);
+}
+
+void s16_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+ Iterator out(dst, window);
+ const int in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const int in_dim_w = src->info()->dimension(1);
+ const int in_dim_h = src->info()->dimension(2);
+ const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+ // Don't increment in Y and Z 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::DimY, Window::Dimension(0, 0, 0));
+ win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+ Iterator in(src, win_in);
+
+ if(border_mode == BorderMode::CONSTANT)
+ {
+ const int16_t const_border_value = static_cast<int16_t>(constant_border_value.get<int16_t>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+ const int16_t *in_ptr = reinterpret_cast<const int16_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+ const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+ const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+ const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+ const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+ *reinterpret_cast<int16_t *>(out.ptr()) = static_cast<int16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else if(border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+ auto clamped_w = utility::clamp<int>(offset, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+ const auto a01 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+ const auto a10 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+ const auto a11 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+ *reinterpret_cast<int16_t *>(out.ptr()) = static_cast<int16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+}
+namespace cpu
+{
+void u8_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ if(policy == InterpolationPolicy::BILINEAR)
+ {
+ u8_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+ }
+ else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ u8_neon_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+ }
+}
+
+void s16_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ if(policy == InterpolationPolicy::BILINEAR)
+ {
+ s16_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+ }
+ else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ s16_neon_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+ }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/scale/neon/list.h b/src/core/cpu/kernels/scale/neon/list.h
new file mode 100644
index 0000000..c91242f
--- /dev/null
+++ b/src/core/cpu/kernels/scale/neon/list.h
@@ -0,0 +1,185 @@
+/*
+ * Copyright (c) 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 SRC_CORE_NEON_KERNELS_SCALE_LIST_H
+#define SRC_CORE_NEON_KERNELS_SCALE_LIST_H
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+#define DECLARE_SCALE_KERNEL(func_name) \
+ void func_name(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, \
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, \
+ bool align_corners, const Window &window)
+
+DECLARE_SCALE_KERNEL(qasymm8_neon_scale);
+DECLARE_SCALE_KERNEL(qasymm8_signed_neon_scale);
+
+#undef DECLARE_SCALE_KERNEL
+
+template <typename T>
+void nearest_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+ const size_t in_stride_wc = in_stride_w * in_stride_c;
+ const size_t in_dim_h = src->info()->dimension(2);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+ 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 int window_step_x = 16 / sizeof(T);
+
+ Window win(window);
+ win.set(Window::DimX, Window::Dimension(0, 1, 1));
+ Iterator out(dst, win);
+
+ const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
+ const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+ execute_window_loop(win, [&](const Coordinates & id)
+ {
+ const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+ const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+ const int offset_row = in_hi * in_stride_wc;
+ int32_t x = window_start_x;
+ const T *in_ptr = reinterpret_cast<const T *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+
+ for(; x <= window_end_x - window_step_x; x += window_step_x)
+ {
+ wrapper::vstore(reinterpret_cast<T *>(out.ptr()) + x,
+ wrapper::vloadq(in_ptr + offset + offset_row + x));
+ }
+ for(; x < window_end_x; ++x)
+ {
+ *(reinterpret_cast<T *>(out.ptr()) + x) = *(in_ptr + offset + offset_row + x);
+ }
+ },
+ out);
+}
+
+template <typename T>
+void bilinear_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+ Iterator out(dst, window);
+ const int in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const int in_dim_w = src->info()->dimension(1);
+ const int in_dim_h = src->info()->dimension(2);
+ const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+ // Don't increment in Y and Z 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::DimY, Window::Dimension(0, 0, 0));
+ win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+ Iterator in(src, win_in);
+
+ if(border_mode == BorderMode::CONSTANT)
+ {
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+ using ConstType = typename std::conditional<std::is_same<T, float16_t>::value, half, T>::type;
+#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+ using ConstType = T;
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+ const T const_border_value = static_cast<T>(constant_border_value.get<ConstType>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+ const T *in_ptr = reinterpret_cast<const T *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+ const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+ const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+ const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+ const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+ *reinterpret_cast<T *>(out.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else if(border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+ auto clamped_w = utility::clamp<int>(offset, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(reinterpret_cast<const T *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+ const auto a01 = *(reinterpret_cast<const T *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+ const auto a10 = *(reinterpret_cast<const T *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+ const auto a11 = *(reinterpret_cast<const T *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+ *reinterpret_cast<T *>(out.ptr()) = static_cast<T>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+
+template <typename T>
+void common_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ if(policy == InterpolationPolicy::BILINEAR)
+ {
+ bilinear_neon_scale<T>(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+ }
+ else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ nearest_neon_scale<T>(src, dst, offsets, sampling_offset, align_corners, window);
+ }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* SRC_CORE_NEON_KERNELS_SCALE_LIST_H */
diff --git a/src/core/cpu/kernels/scale/neon/qasymm8.cpp b/src/core/cpu/kernels/scale/neon/qasymm8.cpp
new file mode 100644
index 0000000..90302ce
--- /dev/null
+++ b/src/core/cpu/kernels/scale/neon/qasymm8.cpp
@@ -0,0 +1,145 @@
+/*
+ * Copyright (c) 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/cpu/kernels/scale/neon/list.h"
+
+namespace arm_compute
+{
+namespace
+{
+void qasymm8_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ // Data layout is NHWC
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+ Window win_off;
+ win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
+ win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+ // 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(1, Window::Dimension(0, 0, 0));
+ win_in.set(2, Window::Dimension(0, 0, 0));
+
+ for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
+ {
+ win_off.set(d, Window::Dimension(0, 0, 0));
+ }
+
+ Iterator in(src, win_in);
+ Iterator out(dst, window);
+
+ const int32_t in_dim_w = src->info()->dimension(1);
+ const int32_t in_dim_h = src->info()->dimension(2);
+ const int32_t stride_w = src->info()->strides_in_bytes()[1];
+ const int32_t stride_h = src->info()->strides_in_bytes()[2];
+
+ const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
+ const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
+
+ if(border_mode == BorderMode::CONSTANT)
+ {
+ const uint8_t const_border_value = static_cast<uint8_t>(constant_border_value.get<uint8_t>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int32_t index_h = std::floor((id[2] + sampling_offset) * hr - sampling_offset);
+ const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[1], id[2]))));
+ const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[1], id[2]))));
+ const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[1], id[2]))));
+ const auto pixel_row_ptr = reinterpret_cast<const uint8_t *>(in.ptr());
+
+ const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ?
+ (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
+ const_border_value;
+ const auto a01 = (-1 <= index_w && index_w < in_dim_w - 1 && 0 <= index_h && index_h < in_dim_h) ?
+ (*(pixel_row_ptr + (index_w + 1) * stride_w + index_h * stride_h)) :
+ const_border_value;
+ const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ?
+ (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) :
+ const_border_value;
+ const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ?
+ (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
+ const_border_value;
+
+ const float inp00 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a00, iq_info);
+ const float inp01 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a01, iq_info);
+ const float inp10 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a10, iq_info);
+ const float inp11 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a11, iq_info);
+ *reinterpret_cast<uint8_t *>(out.ptr()) = Qasymm8QuantizationHelper<uint8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+ },
+ in, out);
+ }
+ else if(border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int index_h = std::floor((id[2] + sampling_offset) * hr - sampling_offset);
+ const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[1], id[2]))));
+ const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[1], id[2]))));
+ const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[1], id[2]))));
+ const auto pixel_row_ptr = reinterpret_cast<const uint8_t *>(in.ptr());
+
+ auto clamped_w = utility::clamp<int>(index_w, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(index_h, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h);
+ const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h);
+ const auto a10 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h1 * stride_h);
+ const auto a11 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h1 * stride_h);
+
+ const float inp00 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a00, iq_info);
+ const float inp01 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a01, iq_info);
+ const float inp10 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a10, iq_info);
+ const float inp11 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a11, iq_info);
+ *reinterpret_cast<uint8_t *>(out.ptr()) = Qasymm8QuantizationHelper<uint8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+ },
+ in, out);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+}
+namespace cpu
+{
+void qasymm8_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ if(policy == InterpolationPolicy::BILINEAR)
+ {
+ qasymm8_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+ }
+ else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ nearest_neon_scale<uint8_t>(src, dst, offsets, sampling_offset, align_corners, window);
+ }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/scale/neon/qasymm8_signed.cpp b/src/core/cpu/kernels/scale/neon/qasymm8_signed.cpp
new file mode 100644
index 0000000..07d6c6e
--- /dev/null
+++ b/src/core/cpu/kernels/scale/neon/qasymm8_signed.cpp
@@ -0,0 +1,145 @@
+/*
+ * Copyright (c) 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/cpu/kernels/scale/neon/list.h"
+
+namespace arm_compute
+{
+namespace
+{
+void qasymm8_signed_neon_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ // Data layout is NHWC
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+ Window win_off;
+ win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
+ win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+ // 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(1, Window::Dimension(0, 0, 0));
+ win_in.set(2, Window::Dimension(0, 0, 0));
+
+ for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
+ {
+ win_off.set(d, Window::Dimension(0, 0, 0));
+ }
+
+ Iterator in(src, win_in);
+ Iterator out(dst, window);
+
+ const int32_t in_dim_w = src->info()->dimension(1);
+ const int32_t in_dim_h = src->info()->dimension(2);
+ const int32_t stride_w = src->info()->strides_in_bytes()[1];
+ const int32_t stride_h = src->info()->strides_in_bytes()[2];
+
+ const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
+ const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
+
+ if(border_mode == BorderMode::CONSTANT)
+ {
+ const int8_t const_border_value = static_cast<int8_t>(constant_border_value.get<int8_t>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int32_t index_h = std::floor((id[2] + sampling_offset) * hr - sampling_offset);
+ const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[1], id[2]))));
+ const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[1], id[2]))));
+ const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[1], id[2]))));
+ const auto pixel_row_ptr = reinterpret_cast<const int8_t *>(in.ptr());
+
+ const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ?
+ (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
+ const_border_value;
+ const auto a01 = (-1 <= index_w && index_w < in_dim_w - 1 && 0 <= index_h && index_h < in_dim_h) ?
+ (*(pixel_row_ptr + (index_w + 1) * stride_w + index_h * stride_h)) :
+ const_border_value;
+ const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ?
+ (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) :
+ const_border_value;
+ const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ?
+ (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
+ const_border_value;
+
+ const float inp00 = Qasymm8QuantizationHelper<int8_t>::dequantize(a00, iq_info);
+ const float inp01 = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info);
+ const float inp10 = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info);
+ const float inp11 = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info);
+ *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+ },
+ in, out);
+ }
+ else if(border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int index_h = std::floor((id[2] + sampling_offset) * hr - sampling_offset);
+ const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[1], id[2]))));
+ const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[1], id[2]))));
+ const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[1], id[2]))));
+ const auto pixel_row_ptr = reinterpret_cast<const int8_t *>(in.ptr());
+
+ auto clamped_w = utility::clamp<int>(index_w, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(index_h, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h);
+ const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h);
+ const auto a10 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h1 * stride_h);
+ const auto a11 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h1 * stride_h);
+
+ const float inp00 = Qasymm8QuantizationHelper<int8_t>::dequantize(a00, iq_info);
+ const float inp01 = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info);
+ const float inp10 = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info);
+ const float inp11 = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info);
+ *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+ },
+ in, out);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+}
+namespace cpu
+{
+void qasymm8_signed_neon_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ if(policy == InterpolationPolicy::BILINEAR)
+ {
+ qasymm8_signed_neon_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+ }
+ else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ nearest_neon_scale<int8_t>(src, dst, offsets, sampling_offset, align_corners, window);
+ }
+}
+} // namespace cpu
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/cpu/kernels/scale/sve/fp16.cpp b/src/core/cpu/kernels/scale/sve/fp16.cpp
new file mode 100644
index 0000000..99f08db
--- /dev/null
+++ b/src/core/cpu/kernels/scale/sve/fp16.cpp
@@ -0,0 +1,176 @@
+/*
+ * Copyright (c) 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace
+{
+void fp16_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+ float sampling_offset, bool align_corners, const Window &window)
+{
+ const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+ const size_t in_stride_wc = in_stride_w * in_stride_c;
+ const size_t in_dim_h = src->info()->dimension(2);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+ const auto window_start_x = static_cast<int32_t>(window.x().start());
+ const auto window_end_x = static_cast<int32_t>(window.x().end());
+
+ Window win(window);
+ win.set(Window::DimX, Window::Dimension(0, 1, 1));
+ Iterator out(dst, win);
+
+ const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
+ const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+ execute_window_loop(win, [&](const Coordinates & id)
+ {
+ const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+ const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+ const int offset_row = in_hi * in_stride_wc;
+ const auto in_ptr = reinterpret_cast<const float16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+ const auto out_ptr = reinterpret_cast<float16_t *>(out.ptr());
+
+ // Compute S elements per iteration
+ int x = window_start_x;
+ svbool_t pg = svwhilelt_b16(x, window_end_x);
+ do
+ {
+ // Store results
+ svst1_f16(pg, out_ptr + x, svld1_f16(pg, in_ptr + offset + offset_row + x));
+
+ x += svcntw();
+ pg = svwhilelt_b16(x, window_end_x);
+ }
+ while(svptest_any(svptrue_b16(), pg));
+ },
+ out);
+}
+
+void fp16_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+ Iterator out(dst, window);
+ const int in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const int in_dim_w = src->info()->dimension(1);
+ const int in_dim_h = src->info()->dimension(2);
+ const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+ // Don't increment in Y and Z 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::DimY, Window::Dimension(0, 0, 0));
+ win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+ Iterator in(src, win_in);
+
+ if(border_mode == BorderMode::CONSTANT)
+ {
+ using ConstType = typename std::conditional<std::is_same<float16_t, float16_t>::value, half, float16_t>::type;
+
+ const float16_t const_border_value = static_cast<float16_t>(constant_border_value.get<ConstType>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+ const float16_t *in_ptr = reinterpret_cast<const float16_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+ const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+ const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+ const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+ const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+ *reinterpret_cast<float16_t *>(out.ptr()) = static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else if(border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+ auto clamped_w = utility::clamp<int>(offset, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+ const auto a01 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+ const auto a10 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+ const auto a11 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+ *reinterpret_cast<float16_t *>(out.ptr()) = static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+}
+namespace cpu
+{
+void fp16_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ if(policy == InterpolationPolicy::BILINEAR)
+ {
+ fp16_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+ }
+ else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ fp16_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+ }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // __ARM_FEATURE_SVE
\ No newline at end of file
diff --git a/src/core/cpu/kernels/scale/sve/fp32.cpp b/src/core/cpu/kernels/scale/sve/fp32.cpp
new file mode 100644
index 0000000..94055ae
--- /dev/null
+++ b/src/core/cpu/kernels/scale/sve/fp32.cpp
@@ -0,0 +1,174 @@
+/*
+ * Copyright (c) 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace
+{
+void fp32_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+ float sampling_offset, bool align_corners, const Window &window)
+{
+ const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+ const size_t in_stride_wc = in_stride_w * in_stride_c;
+ const size_t in_dim_h = src->info()->dimension(2);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+ const auto window_start_x = static_cast<int32_t>(window.x().start());
+ const auto window_end_x = static_cast<int32_t>(window.x().end());
+
+ Window win(window);
+ win.set(Window::DimX, Window::Dimension(0, 1, 1));
+ Iterator out(dst, win);
+
+ const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
+ const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+ execute_window_loop(win, [&](const Coordinates & id)
+ {
+ const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+ const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+ const int offset_row = in_hi * in_stride_wc;
+ const auto in_ptr = reinterpret_cast<const float *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+ const auto out_ptr = reinterpret_cast<float *>(out.ptr());
+
+ // Compute S elements per iteration
+ int x = window_start_x;
+ svbool_t pg = svwhilelt_b32(x, window_end_x);
+ do
+ {
+ // Store results
+ svst1_f32(pg, out_ptr + x, svld1_f32(pg, in_ptr + offset + offset_row + x));
+
+ x += svcntw();
+ pg = svwhilelt_b32(x, window_end_x);
+ }
+ while(svptest_any(svptrue_b32(), pg));
+ },
+ out);
+}
+
+void fp32_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+ Iterator out(dst, window);
+ const int in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const int in_dim_w = src->info()->dimension(1);
+ const int in_dim_h = src->info()->dimension(2);
+ const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+ // Don't increment in Y and Z 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::DimY, Window::Dimension(0, 0, 0));
+ win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+ Iterator in(src, win_in);
+
+ if(border_mode == BorderMode::CONSTANT)
+ {
+ const float const_border_value = static_cast<float>(constant_border_value.get<float>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+ const float *in_ptr = reinterpret_cast<const float *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+ const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+ const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+ const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+ const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+ *reinterpret_cast<float *>(out.ptr()) = static_cast<float>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else if(border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+ auto clamped_w = utility::clamp<int>(offset, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(reinterpret_cast<const float *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+ const auto a01 = *(reinterpret_cast<const float *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+ const auto a10 = *(reinterpret_cast<const float *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+ const auto a11 = *(reinterpret_cast<const float *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+ *reinterpret_cast<float *>(out.ptr()) = static_cast<float>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+}
+namespace cpu
+{
+void fp32_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ if(policy == InterpolationPolicy::BILINEAR)
+ {
+ fp32_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+ }
+ else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ fp32_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+ }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // __ARM_FEATURE_SVE
\ No newline at end of file
diff --git a/src/core/cpu/kernels/scale/sve/integer.cpp b/src/core/cpu/kernels/scale/sve/integer.cpp
new file mode 100644
index 0000000..2a724ec
--- /dev/null
+++ b/src/core/cpu/kernels/scale/sve/integer.cpp
@@ -0,0 +1,301 @@
+/*
+ * Copyright (c) 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace
+{
+void u8_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+ float sampling_offset, bool align_corners, const Window &window)
+{
+ const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+ const size_t in_stride_wc = in_stride_w * in_stride_c;
+ const size_t in_dim_h = src->info()->dimension(2);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+ const auto window_start_x = static_cast<int32_t>(window.x().start());
+ const auto window_end_x = static_cast<int32_t>(window.x().end());
+
+ Window win(window);
+ win.set(Window::DimX, Window::Dimension(0, 1, 1));
+ Iterator out(dst, win);
+
+ const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
+ const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+ execute_window_loop(win, [&](const Coordinates & id)
+ {
+ const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+ const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+ const int offset_row = in_hi * in_stride_wc;
+ const auto in_ptr = reinterpret_cast<const uint8_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+ const auto out_ptr = reinterpret_cast<uint8_t *>(out.ptr());
+
+ // Compute S elements per iteration
+ int x = window_start_x;
+ svbool_t pg = svwhilelt_b8(x, window_end_x);
+ do
+ {
+ // Store results
+ svst1_u8(pg, out_ptr + x, svld1_u8(pg, in_ptr + offset + offset_row + x));
+
+ x += svcntw();
+ pg = svwhilelt_b8(x, window_end_x);
+ }
+ while(svptest_any(svptrue_b8(), pg));
+ },
+ out);
+}
+
+void u8_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+ Iterator out(dst, window);
+ const int in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const int in_dim_w = src->info()->dimension(1);
+ const int in_dim_h = src->info()->dimension(2);
+ const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+ // Don't increment in Y and Z 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::DimY, Window::Dimension(0, 0, 0));
+ win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+ Iterator in(src, win_in);
+
+ if(border_mode == BorderMode::CONSTANT)
+ {
+ const uint8_t const_border_value = static_cast<uint8_t>(constant_border_value.get<uint8_t>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+ const uint8_t *in_ptr = reinterpret_cast<const uint8_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+ const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+ const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+ const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+ const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+ *reinterpret_cast<uint8_t *>(out.ptr()) = static_cast<uint8_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else if(border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+ auto clamped_w = utility::clamp<int>(offset, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+ const auto a01 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+ const auto a10 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+ const auto a11 = *(reinterpret_cast<const uint8_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+ *reinterpret_cast<uint8_t *>(out.ptr()) = static_cast<uint8_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+
+void s16_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+ float sampling_offset, bool align_corners, const Window &window)
+{
+ const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+ const size_t in_stride_wc = in_stride_w * in_stride_c;
+ const size_t in_dim_h = src->info()->dimension(2);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+ const auto window_start_x = static_cast<int32_t>(window.x().start());
+ const auto window_end_x = static_cast<int32_t>(window.x().end());
+
+ Window win(window);
+ win.set(Window::DimX, Window::Dimension(0, 1, 1));
+ Iterator out(dst, win);
+
+ const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
+ const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+ execute_window_loop(win, [&](const Coordinates & id)
+ {
+ const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+ const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+ const int offset_row = in_hi * in_stride_wc;
+ const auto in_ptr = reinterpret_cast<const int16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+ const auto out_ptr = reinterpret_cast<int16_t *>(out.ptr());
+
+ // Compute S elements per iteration
+ int x = window_start_x;
+ svbool_t pg = svwhilelt_b16(x, window_end_x);
+ do
+ {
+ // Store results
+ svst1_s16(pg, out_ptr + x, svld1_s16(pg, in_ptr + offset + offset_row + x));
+
+ x += svcntw();
+ pg = svwhilelt_b16(x, window_end_x);
+ }
+ while(svptest_any(svptrue_b16(), pg));
+ },
+ out);
+}
+
+void s16_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners);
+
+ Iterator out(dst, window);
+ const int in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const int in_dim_w = src->info()->dimension(1);
+ const int in_dim_h = src->info()->dimension(2);
+ const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom);
+
+ // Don't increment in Y and Z 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::DimY, Window::Dimension(0, 0, 0));
+ win_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
+ Iterator in(src, win_in);
+
+ if(border_mode == BorderMode::CONSTANT)
+ {
+ const int16_t const_border_value = static_cast<int16_t>(constant_border_value.get<int16_t>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+ const int16_t *in_ptr = reinterpret_cast<const int16_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc;
+
+ const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value;
+ const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value;
+ const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value;
+ const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value;
+
+ *reinterpret_cast<int16_t *>(out.ptr()) = static_cast<int16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else if(border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z())));
+ const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z())));
+ const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset);
+
+ auto clamped_w = utility::clamp<int>(offset, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(in_hi, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc);
+ const auto a01 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc);
+ const auto a10 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc);
+ const auto a11 = *(reinterpret_cast<const int16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc);
+
+ *reinterpret_cast<int16_t *>(out.ptr()) = static_cast<int16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val));
+ },
+ in, out);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+}
+namespace cpu
+{
+void u8_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ if(policy == InterpolationPolicy::BILINEAR)
+ {
+ u8_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+ }
+ else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ u8_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+ }
+}
+
+void s16_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ if(policy == InterpolationPolicy::BILINEAR)
+ {
+ s16_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+ }
+ else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ s16_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+ }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // __ARM_FEATURE_SVE
\ No newline at end of file
diff --git a/src/core/cpu/kernels/scale/sve/list.h b/src/core/cpu/kernels/scale/sve/list.h
new file mode 100644
index 0000000..b9c3a10
--- /dev/null
+++ b/src/core/cpu/kernels/scale/sve/list.h
@@ -0,0 +1,47 @@
+/*
+ * Copyright (c) 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 SRC_CORE_SVE_KERNELS_SCALE_LIST_H
+#define SRC_CORE_SVE_KERNELS_SCALE_LIST_H
+
+namespace arm_compute
+{
+namespace cpu
+{
+#define DECLARE_SCALE_KERNEL(func_name) \
+ void func_name(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, \
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, \
+ bool align_corners, const Window &window)
+
+DECLARE_SCALE_KERNEL(fp16_sve_scale);
+DECLARE_SCALE_KERNEL(fp32_sve_scale);
+DECLARE_SCALE_KERNEL(s16_sve_scale);
+DECLARE_SCALE_KERNEL(u8_sve_scale);
+DECLARE_SCALE_KERNEL(qasymm8_sve_scale);
+DECLARE_SCALE_KERNEL(qasymm8_signed_sve_scale);
+
+#undef DECLARE_SCALE_KERNEL
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* SRC_CORE_SVE_KERNELS_SCALE_LIST_H */
diff --git a/src/core/cpu/kernels/scale/sve/qasymm8.cpp b/src/core/cpu/kernels/scale/sve/qasymm8.cpp
new file mode 100644
index 0000000..c475ad6
--- /dev/null
+++ b/src/core/cpu/kernels/scale/sve/qasymm8.cpp
@@ -0,0 +1,209 @@
+/*
+ * Copyright (c) 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace
+{
+void qasymm8_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+ float sampling_offset, bool align_corners, const Window &window)
+{
+ const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+ const size_t in_stride_wc = in_stride_w * in_stride_c;
+ const size_t in_dim_h = src->info()->dimension(2);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+ const auto window_start_x = static_cast<int32_t>(window.x().start());
+ const auto window_end_x = static_cast<int32_t>(window.x().end());
+
+ Window win(window);
+ win.set(Window::DimX, Window::Dimension(0, 1, 1));
+ Iterator out(dst, win);
+
+ const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
+ const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+ execute_window_loop(win, [&](const Coordinates & id)
+ {
+ const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+ const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+ const int offset_row = in_hi * in_stride_wc;
+ const auto in_ptr = reinterpret_cast<const uint8_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+ const auto out_ptr = reinterpret_cast<uint8_t *>(out.ptr());
+
+ // Compute S elements per iteration
+ int x = window_start_x;
+ svbool_t pg = svwhilelt_b8(x, window_end_x);
+ do
+ {
+ // Store results
+ svst1_u8(pg, out_ptr + x, svld1_u8(pg, in_ptr + offset + offset_row + x));
+
+ x += svcntw();
+ pg = svwhilelt_b8(x, window_end_x);
+ }
+ while(svptest_any(svptrue_b8(), pg));
+ },
+ out);
+}
+
+void qasymm8_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ // Get data layout and width/height indices
+ const DataLayout data_layout = src->info()->data_layout();
+ const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+ const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(idx_height), dst->info()->dimension(idx_height), align_corners);
+ Window win_off;
+ win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
+ win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+ // 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(idx_width, Window::Dimension(0, 0, 0));
+ win_in.set(idx_height, Window::Dimension(0, 0, 0));
+
+ for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
+ {
+ win_off.set(d, Window::Dimension(0, 0, 0));
+ }
+
+ Iterator in(src, win_in);
+ Iterator out(dst, window);
+
+ const int32_t in_dim_w = src->info()->dimension(idx_width);
+ const int32_t in_dim_h = src->info()->dimension(idx_height);
+ const int32_t stride_w = src->info()->strides_in_bytes()[idx_width];
+ const int32_t stride_h = src->info()->strides_in_bytes()[idx_height];
+
+ const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
+ const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
+
+ if(border_mode == BorderMode::CONSTANT)
+ {
+ const uint8_t const_border_value = static_cast<uint8_t>(constant_border_value.get<uint8_t>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int32_t index_h = std::floor((id[idx_height] + sampling_offset) * hr - sampling_offset);
+ const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto pixel_row_ptr = reinterpret_cast<const uint8_t *>(in.ptr());
+
+ const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ?
+ (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
+ const_border_value;
+ const auto a01 = (-1 <= index_w && index_w < in_dim_w - 1 && 0 <= index_h && index_h < in_dim_h) ?
+ (*(pixel_row_ptr + (index_w + 1) * stride_w + index_h * stride_h)) :
+ const_border_value;
+ const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ?
+ (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) :
+ const_border_value;
+ const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ?
+ (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
+ const_border_value;
+
+ const float inp00 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a00, iq_info);
+ const float inp01 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a01, iq_info);
+ const float inp10 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a10, iq_info);
+ const float inp11 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a11, iq_info);
+ *reinterpret_cast<uint8_t *>(out.ptr()) = Qasymm8QuantizationHelper<uint8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+ },
+ in, out);
+ }
+ else if(border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int index_h = std::floor((id[idx_height] + sampling_offset) * hr - sampling_offset);
+ const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto pixel_row_ptr = reinterpret_cast<const uint8_t *>(in.ptr());
+
+ auto clamped_w = utility::clamp<int>(index_w, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(index_h, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h);
+ const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h);
+ const auto a10 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h1 * stride_h);
+ const auto a11 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h1 * stride_h);
+
+ const float inp00 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a00, iq_info);
+ const float inp01 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a01, iq_info);
+ const float inp10 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a10, iq_info);
+ const float inp11 = Qasymm8QuantizationHelper<uint8_t>::dequantize(a11, iq_info);
+ *reinterpret_cast<uint8_t *>(out.ptr()) = Qasymm8QuantizationHelper<uint8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+ },
+ in, out);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+}
+namespace cpu
+{
+void qasymm8_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ if(policy == InterpolationPolicy::BILINEAR)
+ {
+ qasymm8_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+ }
+ else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ qasymm8_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+ }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // __ARM_FEATURE_SVE
\ No newline at end of file
diff --git a/src/core/cpu/kernels/scale/sve/qasymm8_signed.cpp b/src/core/cpu/kernels/scale/sve/qasymm8_signed.cpp
new file mode 100644
index 0000000..b39b75a
--- /dev/null
+++ b/src/core/cpu/kernels/scale/sve/qasymm8_signed.cpp
@@ -0,0 +1,209 @@
+/*
+ * Copyright (c) 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 "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/helpers/ScaleHelpers.h"
+#include "src/core/utils/ScaleUtils.h"
+#include "support/Rounding.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace
+{
+void qasymm8_signed_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets,
+ float sampling_offset, bool align_corners, const Window &window)
+{
+ const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right;
+ const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom;
+ const size_t in_stride_wc = in_stride_w * in_stride_c;
+ const size_t in_dim_h = src->info()->dimension(2);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners);
+ const auto window_start_x = static_cast<int32_t>(window.x().start());
+ const auto window_end_x = static_cast<int32_t>(window.x().end());
+
+ Window win(window);
+ win.set(Window::DimX, Window::Dimension(0, 1, 1));
+ Iterator out(dst, win);
+
+ const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes();
+ const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3];
+
+ execute_window_loop(win, [&](const Coordinates & id)
+ {
+ const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c;
+ const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr));
+ const int offset_row = in_hi * in_stride_wc;
+ const auto in_ptr = reinterpret_cast<const int8_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]);
+ const auto out_ptr = reinterpret_cast<int8_t *>(out.ptr());
+
+ // Compute S elements per iteration
+ int x = window_start_x;
+ svbool_t pg = svwhilelt_b8(x, window_end_x);
+ do
+ {
+ // Store results
+ svst1_s8(pg, out_ptr + x, svld1_s8(pg, in_ptr + offset + offset_row + x));
+
+ x += svcntw();
+ pg = svwhilelt_b8(x, window_end_x);
+ }
+ while(svptest_any(svptrue_b8(), pg));
+ },
+ out);
+}
+
+void qasymm8_signed_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ // Get data layout and width/height indices
+ const DataLayout data_layout = src->info()->data_layout();
+ const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+ const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+
+ // Compute the ratio between source height and destination height
+ const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(idx_height), dst->info()->dimension(idx_height), align_corners);
+ Window win_off;
+ win_off.set(Window::DimX, Window::Dimension(0, 0, 0));
+ win_off.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+ // 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(idx_width, Window::Dimension(0, 0, 0));
+ win_in.set(idx_height, Window::Dimension(0, 0, 0));
+
+ for(size_t d = Window::DimZ; d < offsets->info()->num_dimensions(); ++d)
+ {
+ win_off.set(d, Window::Dimension(0, 0, 0));
+ }
+
+ Iterator in(src, win_in);
+ Iterator out(dst, window);
+
+ const int32_t in_dim_w = src->info()->dimension(idx_width);
+ const int32_t in_dim_h = src->info()->dimension(idx_height);
+ const int32_t stride_w = src->info()->strides_in_bytes()[idx_width];
+ const int32_t stride_h = src->info()->strides_in_bytes()[idx_height];
+
+ const UniformQuantizationInfo iq_info = src->info()->quantization_info().uniform();
+ const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
+
+ if(border_mode == BorderMode::CONSTANT)
+ {
+ const int8_t const_border_value = static_cast<int8_t>(constant_border_value.get<int8_t>());
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int32_t index_h = std::floor((id[idx_height] + sampling_offset) * hr - sampling_offset);
+ const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto pixel_row_ptr = reinterpret_cast<const int8_t *>(in.ptr());
+
+ const auto a00 = (0 <= index_w && index_w < in_dim_w && 0 <= index_h && index_h < in_dim_h) ?
+ (*(pixel_row_ptr + index_w * stride_w + index_h * stride_h)) :
+ const_border_value;
+ const auto a01 = (-1 <= index_w && index_w < in_dim_w - 1 && 0 <= index_h && index_h < in_dim_h) ?
+ (*(pixel_row_ptr + (index_w + 1) * stride_w + index_h * stride_h)) :
+ const_border_value;
+ const auto a10 = (0 <= index_w && index_w < in_dim_w && -1 <= index_h && index_h < in_dim_h - 1) ?
+ (*(pixel_row_ptr + index_w * stride_w + (index_h + 1) * stride_h)) :
+ const_border_value;
+ const auto a11 = (-1 <= index_w && index_w < in_dim_w - 1 && -1 <= index_h && index_h < in_dim_h - 1) ?
+ (*(pixel_row_ptr + (index_w + 1) * stride_w + (index_h + 1) * stride_h)) :
+ const_border_value;
+
+ const float inp00 = Qasymm8QuantizationHelper<int8_t>::dequantize(a00, iq_info);
+ const float inp01 = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info);
+ const float inp10 = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info);
+ const float inp11 = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info);
+ *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+ },
+ in, out);
+ }
+ else if(border_mode == BorderMode::REPLICATE)
+ {
+ execute_window_loop(window, [&](const Coordinates & id)
+ {
+ const int index_h = std::floor((id[idx_height] + sampling_offset) * hr - sampling_offset);
+ const int32_t index_w = *(reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto dx_val = *(reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto dy_val = *(reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id[idx_width], id[idx_height]))));
+ const auto pixel_row_ptr = reinterpret_cast<const int8_t *>(in.ptr());
+
+ auto clamped_w = utility::clamp<int>(index_w, 0, in_dim_w - 1);
+ auto clamped_w1 = utility::clamp<int>(index_w + 1, 0, in_dim_w - 1);
+ auto clamped_h = utility::clamp<int>(index_h, 0, in_dim_h - 1);
+ auto clamped_h1 = utility::clamp<int>(index_h + 1, 0, in_dim_h - 1);
+
+ const auto a00 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h * stride_h);
+ const auto a01 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h * stride_h);
+ const auto a10 = *(pixel_row_ptr + clamped_w * stride_w + clamped_h1 * stride_h);
+ const auto a11 = *(pixel_row_ptr + clamped_w1 * stride_w + clamped_h1 * stride_h);
+
+ const float inp00 = Qasymm8QuantizationHelper<int8_t>::dequantize(a00, iq_info);
+ const float inp01 = Qasymm8QuantizationHelper<int8_t>::dequantize(a01, iq_info);
+ const float inp10 = Qasymm8QuantizationHelper<int8_t>::dequantize(a10, iq_info);
+ const float inp11 = Qasymm8QuantizationHelper<int8_t>::dequantize(a11, iq_info);
+ *reinterpret_cast<int8_t *>(out.ptr()) = Qasymm8QuantizationHelper<int8_t>::quantize(scale_helpers::delta_bilinear(inp00, inp01, inp10, inp11, dx_val, dy_val), oq_info);
+ },
+ in, out);
+ }
+ else
+ {
+ ARM_COMPUTE_ERROR("Not implemented");
+ }
+}
+}
+namespace cpu
+{
+void qasymm8_signed_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy,
+ InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset,
+ bool align_corners, const Window &window)
+{
+ if(policy == InterpolationPolicy::BILINEAR)
+ {
+ qasymm8_signed_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window);
+ }
+ else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR)
+ {
+ qasymm8_signed_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window);
+ }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // __ARM_FEATURE_SVE
\ No newline at end of file