aboutsummaryrefslogtreecommitdiff
path: root/src/cpu/kernels/scale
diff options
context:
space:
mode:
Diffstat (limited to 'src/cpu/kernels/scale')
-rw-r--r--src/cpu/kernels/scale/neon/fp16.cpp174
-rw-r--r--src/cpu/kernels/scale/neon/integer.cpp293
-rw-r--r--src/cpu/kernels/scale/neon/list.h185
-rw-r--r--src/cpu/kernels/scale/neon/qasymm8.cpp145
-rw-r--r--src/cpu/kernels/scale/neon/qasymm8_signed.cpp145
-rw-r--r--src/cpu/kernels/scale/sve/fp16.cpp176
-rw-r--r--src/cpu/kernels/scale/sve/fp32.cpp174
-rw-r--r--src/cpu/kernels/scale/sve/integer.cpp300
-rw-r--r--src/cpu/kernels/scale/sve/list.h47
-rw-r--r--src/cpu/kernels/scale/sve/qasymm8.cpp207
-rw-r--r--src/cpu/kernels/scale/sve/qasymm8_signed.cpp207
11 files changed, 2053 insertions, 0 deletions
diff --git a/src/cpu/kernels/scale/neon/fp16.cpp b/src/cpu/kernels/scale/neon/fp16.cpp
new file mode 100644
index 0000000000..0ad66cab1c
--- /dev/null
+++ b/src/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/cpu/kernels/scale/neon/integer.cpp b/src/cpu/kernels/scale/neon/integer.cpp
new file mode 100644
index 0000000000..a2359aac94
--- /dev/null
+++ b/src/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/cpu/kernels/scale/neon/list.h b/src/cpu/kernels/scale/neon/list.h
new file mode 100644
index 0000000000..c91242f5b2
--- /dev/null
+++ b/src/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/cpu/kernels/scale/neon/qasymm8.cpp b/src/cpu/kernels/scale/neon/qasymm8.cpp
new file mode 100644
index 0000000000..fb52752690
--- /dev/null
+++ b/src/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/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/cpu/kernels/scale/neon/qasymm8_signed.cpp b/src/cpu/kernels/scale/neon/qasymm8_signed.cpp
new file mode 100644
index 0000000000..706bcee954
--- /dev/null
+++ b/src/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/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/cpu/kernels/scale/sve/fp16.cpp b/src/cpu/kernels/scale/sve/fp16.cpp
new file mode 100644
index 0000000000..76e7735b8a
--- /dev/null
+++ b/src/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.
+ */
+
+#if defined(ARM_COMPUTE_ENABLE_SVE)
+#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_sve.h>
+#include <cmath>
+#include <cstddef>
+
+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_COMPUTE_ENABLE_SVE \ No newline at end of file
diff --git a/src/cpu/kernels/scale/sve/fp32.cpp b/src/cpu/kernels/scale/sve/fp32.cpp
new file mode 100644
index 0000000000..030e109cdf
--- /dev/null
+++ b/src/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.
+ */
+#if defined(ARM_COMPUTE_ENABLE_SVE)
+#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>
+
+#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_COMPUTE_ENABLE_SVE \ No newline at end of file
diff --git a/src/cpu/kernels/scale/sve/integer.cpp b/src/cpu/kernels/scale/sve/integer.cpp
new file mode 100644
index 0000000000..486c674612
--- /dev/null
+++ b/src/cpu/kernels/scale/sve/integer.cpp
@@ -0,0 +1,300 @@
+/*
+ * 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.
+ */
+#if defined(ARM_COMPUTE_ENABLE_SVE)
+#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_sve.h>
+#include <cmath>
+#include <cstddef>
+
+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_COMPUTE_ENABLE_SVE \ No newline at end of file
diff --git a/src/cpu/kernels/scale/sve/list.h b/src/cpu/kernels/scale/sve/list.h
new file mode 100644
index 0000000000..b9c3a10a78
--- /dev/null
+++ b/src/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/cpu/kernels/scale/sve/qasymm8.cpp b/src/cpu/kernels/scale/sve/qasymm8.cpp
new file mode 100644
index 0000000000..c9122ad40b
--- /dev/null
+++ b/src/cpu/kernels/scale/sve/qasymm8.cpp
@@ -0,0 +1,207 @@
+/*
+ * 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.
+ */
+#if defined(ARM_COMPUTE_ENABLE_SVE)
+#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 <arm_sve.h>
+#include <cmath>
+#include <cstddef>
+
+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)
+{
+ // Data layout is NHWC
+ const int idx_width = 1;
+ const int idx_height = 2;
+
+ // 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 // defined(ARM_COMPUTE_ENABLE_SVE) \ No newline at end of file
diff --git a/src/cpu/kernels/scale/sve/qasymm8_signed.cpp b/src/cpu/kernels/scale/sve/qasymm8_signed.cpp
new file mode 100644
index 0000000000..0843e61fd4
--- /dev/null
+++ b/src/cpu/kernels/scale/sve/qasymm8_signed.cpp
@@ -0,0 +1,207 @@
+/*
+ * 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.
+ */
+#if defined(ARM_COMPUTE_ENABLE_SVE)
+#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 <arm_sve.h>
+#include <cmath>
+#include <cstddef>
+
+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)
+{
+ // Data layout is NHWC
+ const int idx_width = 1;
+ const int idx_height = 2;
+
+ // 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_COMPUTE_ENABLE_SVE \ No newline at end of file
generated by cgit v1.2.1 (git 2.23.0) at 2024-06-02 22:48:57 +0000