aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/NEReverseKernel.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/NEReverseKernel.cpp')
-rw-r--r--src/core/NEON/kernels/NEReverseKernel.cpp140
1 files changed, 87 insertions, 53 deletions
diff --git a/src/core/NEON/kernels/NEReverseKernel.cpp b/src/core/NEON/kernels/NEReverseKernel.cpp
index 758433f89f..b3710555df 100644
--- a/src/core/NEON/kernels/NEReverseKernel.cpp
+++ b/src/core/NEON/kernels/NEReverseKernel.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018-2021 Arm Limited.
+ * Copyright (c) 2018-2021, 2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -26,25 +26,30 @@
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/Window.h"
-#include "src/core/NEON/wrapper/wrapper.h"
+
#include "src/core/helpers/AutoConfiguration.h"
#include "src/core/helpers/WindowHelpers.h"
+#include "src/core/NEON/wrapper/wrapper.h"
namespace arm_compute
{
namespace
{
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *axis)
+Status
+validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *axis, bool use_inverted_axis)
{
+ ARM_COMPUTE_UNUSED(use_inverted_axis);
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output, axis);
//Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use CPU FP16 instructions.
ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(axis, 1, DataType::U32);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(axis, 1, DataType::U32, DataType::S32);
ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis->num_dimensions() > 1, "Axis must be a 1D tensor");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 4,
+ "Current implementation only supports up to 4 dimensions.");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis->dimension(0) > 4, "Only up to 4 dimensions can be reversed");
// Checks performed when output is configured
- if(output->total_size() != 0)
+ if (output->total_size() != 0)
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
@@ -55,42 +60,67 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c
}
} // namespace
-NEReverseKernel::NEReverseKernel()
- : _input(nullptr), _output(nullptr), _axis(nullptr)
+NEReverseKernel::NEReverseKernel() : _input(nullptr), _output(nullptr), _axis(nullptr), _use_inverted_axis(false)
{
}
-void NEReverseKernel::configure(const ITensor *input, ITensor *output, const ITensor *axis)
+void NEReverseKernel::configure(const ITensor *input, ITensor *output, const ITensor *axis, bool use_inverted_axis)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, axis);
- _input = input;
- _output = output;
- _axis = axis;
+ _input = input;
+ _output = output;
+ _axis = axis;
+ _use_inverted_axis = use_inverted_axis;
// Output tensor auto initialization if not yet initialized
auto_init_if_empty(*output->info(), *input->info()->clone());
- ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), axis->info()));
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), axis->info(), use_inverted_axis));
// Configure kernel window
INEKernel::configure(calculate_max_window(*output->info()));
}
-Status NEReverseKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *axis)
+Status NEReverseKernel::validate(const ITensorInfo *input,
+ const ITensorInfo *output,
+ const ITensorInfo *axis,
+ bool use_inverted_axis)
{
- ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, axis));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, axis, use_inverted_axis));
return Status{};
}
template <typename T>
-void run_reverse(const Window &window, const ITensor *input, const ITensor *axis, ITensor *output)
+void run_reverse(
+ const Window &window, const ITensor *input, const ITensor *axis, ITensor *output, bool use_inverted_axis)
{
- int axis_bit = 0;
- for(unsigned int i = 0; i < axis->info()->dimension(0); ++i)
+ unsigned int axis_bit = 0;
+ const int rank = input->info()->num_dimensions();
+
+ for (unsigned int i = 0; i < axis->info()->dimension(0); ++i)
{
- const int axis_i = *(reinterpret_cast<const int *>(axis->buffer()) + i);
+ int axis_i = *(reinterpret_cast<const int *>(axis->buffer()) + i);
+
+ // The values of axis tensor must be between [-rank, rank-1].
+ if ((axis_i < -rank) || (axis_i >= rank))
+ {
+ ARM_COMPUTE_ERROR("the values of the axis tensor must be within [-rank, rank-1].");
+ }
+
+ // In case of negative axis value i.e targeted axis(i) = rank + axis(i)
+ if (axis_i < 0)
+ {
+ axis_i = rank + axis_i;
+ }
+
+ // Reverse ACL axis indices convention i.e. (inverted)axis = (tensor_rank - 1) - axis
+ if (use_inverted_axis)
+ {
+ axis_i = (rank - 1) - axis_i;
+ }
+
axis_bit |= 1 << axis_i;
}
@@ -103,43 +133,47 @@ void run_reverse(const Window &window, const ITensor *input, const ITensor *axis
win.set(Window::DimX, Window::Dimension(0, 1, 1));
Iterator input_it(input, win);
- execute_window_loop(win, [&](const Coordinates & id)
- {
- int x = window_start_x;
- for(; x <= (window_end_x - window_step_x); x += window_step_x)
+ execute_window_loop(
+ win,
+ [&](const Coordinates &id)
{
- auto in = wrapper::vloadq(reinterpret_cast<T *>(input_it.ptr()) + x);
-
- // Reverse 0 axis
- if(axis_bit & 0x1)
+ int x = window_start_x;
+ for (; x <= (window_end_x - window_step_x); x += window_step_x)
{
- in = wrapper::vrev64(in);
- in = wrapper::vcombine(wrapper::vgethigh(in), wrapper::vgetlow(in));
+ auto in = wrapper::vloadq(reinterpret_cast<T *>(input_it.ptr()) + x);
+
+ // Reverse 0 axis
+ if (axis_bit & 0x1)
+ {
+ in = wrapper::vrev64(in);
+ in = wrapper::vcombine(wrapper::vgethigh(in), wrapper::vgetlow(in));
+ }
+
+ const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - x - window_step_x : x;
+ const int offset_y = (axis_bit & 0x2) ? output->info()->dimension(1) - id.y() - 1 : id.y();
+ const int offset_z = (axis_bit & 0x4) ? output->info()->dimension(2) - id.z() - 1 : id.z();
+ const int offset_w = (axis_bit & 0x8) ? output->info()->dimension(3) - id[3] - 1 : id[3];
+
+ auto out_ptr =
+ reinterpret_cast<T *>(output->ptr_to_element(Coordinates(offset_x, offset_y, offset_z, offset_w)));
+ wrapper::vstore(out_ptr, in);
}
- const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - x - window_step_x : x;
- const int offset_y = (axis_bit & 0x2) ? output->info()->dimension(1) - id.y() - 1 : id.y();
- const int offset_z = (axis_bit & 0x4) ? output->info()->dimension(2) - id.z() - 1 : id.z();
- const int offset_w = (axis_bit & 0x8) ? output->info()->dimension(3) - id[3] - 1 : id[3];
-
- auto out_ptr = reinterpret_cast<T *>(output->ptr_to_element(Coordinates(offset_x, offset_y, offset_z, offset_w)));
- wrapper::vstore(out_ptr, in);
- }
-
- // Compute left-over elements
- for(; x < window_end_x; ++x)
- {
- const auto in = *(reinterpret_cast<T *>(input_it.ptr()) + x);
+ // Compute left-over elements
+ for (; x < window_end_x; ++x)
+ {
+ const auto in = *(reinterpret_cast<T *>(input_it.ptr()) + x);
- const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - x - 1 : x;
- const int offset_y = (axis_bit & 0x2) ? output->info()->dimension(1) - id.y() - 1 : id.y();
- const int offset_z = (axis_bit & 0x4) ? output->info()->dimension(2) - id.z() - 1 : id.z();
- const int offset_w = (axis_bit & 0x8) ? output->info()->dimension(3) - id[3] - 1 : id[3];
+ const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - x - 1 : x;
+ const int offset_y = (axis_bit & 0x2) ? output->info()->dimension(1) - id.y() - 1 : id.y();
+ const int offset_z = (axis_bit & 0x4) ? output->info()->dimension(2) - id.z() - 1 : id.z();
+ const int offset_w = (axis_bit & 0x8) ? output->info()->dimension(3) - id[3] - 1 : id[3];
- *reinterpret_cast<T *>(output->ptr_to_element(Coordinates(offset_x, offset_y, offset_z, offset_w))) = in;
- }
- },
- input_it);
+ *reinterpret_cast<T *>(output->ptr_to_element(Coordinates(offset_x, offset_y, offset_z, offset_w))) =
+ in;
+ }
+ },
+ input_it);
}
void NEReverseKernel::run(const Window &window, const ThreadInfo &info)
@@ -148,16 +182,16 @@ void NEReverseKernel::run(const Window &window, const ThreadInfo &info)
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
- switch(_input->info()->element_size())
+ switch (_input->info()->element_size())
{
case 4:
- run_reverse<uint32_t>(window, _input, _axis, _output);
+ run_reverse<uint32_t>(window, _input, _axis, _output, _use_inverted_axis);
break;
case 2:
- run_reverse<uint16_t>(window, _input, _axis, _output);
+ run_reverse<uint16_t>(window, _input, _axis, _output, _use_inverted_axis);
break;
case 1:
- run_reverse<uint8_t>(window, _input, _axis, _output);
+ run_reverse<uint8_t>(window, _input, _axis, _output, _use_inverted_axis);
break;
default:
ARM_COMPUTE_ERROR("Element size not supported");
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-03 17:18:08 +0000