COMPMID-2246: Implement NEBoundingBoxTransform.

Change-Id: I69faa21ccdc295f5369450a12e6e217d873349e3
Signed-off-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1854
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

2 files changed, 233 insertions, 0 deletions

diff --git a/src/core/NEON/kernels/NEBoundingBoxTransformKernel.cpp b/src/core/NEON/kernels/NEBoundingBoxTransformKernel.cpp
new file mode 100644
index 0000000000..cfd3e708e5
--- /dev/null
+++ b/src/core/NEON/kernels/NEBoundingBoxTransformKernel.cpp
@@ -0,0 +1,191 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/NEON/kernels/NEBoundingBoxTransformKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CPP/Validate.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Window.h"
+
+#include <arm_neon.h>
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *boxes, const ITensorInfo *pred_boxes, const ITensorInfo *deltas, const BoundingBoxTransformInfo &info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(boxes, pred_boxes, deltas);
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(boxes);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(boxes, DataType::F32, DataType::F16);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(deltas, DataType::F32, DataType::F16);
+    ARM_COMPUTE_RETURN_ERROR_ON(deltas->tensor_shape()[1] != boxes->tensor_shape()[1]);
+    ARM_COMPUTE_RETURN_ERROR_ON(deltas->tensor_shape()[0] % 4 != 0);
+    ARM_COMPUTE_RETURN_ERROR_ON(boxes->tensor_shape()[0] != 4);
+    ARM_COMPUTE_RETURN_ERROR_ON(deltas->num_dimensions() > 2);
+    ARM_COMPUTE_RETURN_ERROR_ON(boxes->num_dimensions() > 2);
+
+    if(pred_boxes->total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(pred_boxes->tensor_shape(), deltas->tensor_shape());
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(pred_boxes, deltas);
+        ARM_COMPUTE_RETURN_ERROR_ON(pred_boxes->num_dimensions() > 2);
+    }
+    ARM_COMPUTE_RETURN_ERROR_ON(info.scale() <= 0);
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *boxes, ITensorInfo *pred_boxes, ITensorInfo *deltas, const BoundingBoxTransformInfo &bb_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(boxes, pred_boxes);
+
+    auto_init_if_empty(*pred_boxes, *deltas);
+
+    const unsigned int num_boxes = boxes->dimension(1);
+
+    Window window;
+    window.set(Window::DimX, Window::Dimension(0, 1u));
+    window.set(Window::DimY, Window::Dimension(0, num_boxes));
+
+    AccessWindowHorizontal input_access(boxes, 0, 4u);
+    AccessWindowHorizontal output_access(pred_boxes, 0, 4u);
+
+    const bool window_changed = update_window_and_padding(window, input_access, output_access);
+    output_access.set_valid_region(window, ValidRegion(Coordinates(), pred_boxes->tensor_shape()));
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, window);
+}
+
+} // namespace
+
+NEBoundingBoxTransformKernel::NEBoundingBoxTransformKernel()
+    : _boxes(nullptr), _pred_boxes(nullptr), _deltas(nullptr), _bbinfo(0, 0, 0)
+{
+}
+
+void NEBoundingBoxTransformKernel::configure(const ITensor *boxes, ITensor *pred_boxes, const ITensor *deltas, const BoundingBoxTransformInfo &info)
+{
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(boxes->info(), pred_boxes->info(), deltas->info(), info));
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(boxes->info(), pred_boxes->info(), deltas->info(), info);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+
+    // Set instance variables
+    _boxes      = boxes;
+    _pred_boxes = pred_boxes;
+    _deltas     = deltas;
+    _bbinfo     = info;
+
+    INEKernel::configure(win_config.second);
+}
+
+Status NEBoundingBoxTransformKernel::validate(const ITensorInfo *boxes, const ITensorInfo *pred_boxes, const ITensorInfo *deltas, const BoundingBoxTransformInfo &info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(boxes, pred_boxes, deltas, info));
+    return Status{};
+}
+
+template <typename T>
+void NEBoundingBoxTransformKernel::internal_run(const Window &window, const ThreadInfo &info)
+{
+    const size_t num_classes  = _deltas->info()->tensor_shape()[0] >> 2;
+    const size_t deltas_width = _deltas->info()->tensor_shape()[0];
+    const int    img_h        = floor(_bbinfo.img_height() / _bbinfo.scale() + 0.5f);
+    const int    img_w        = floor(_bbinfo.img_width() / _bbinfo.scale() + 0.5f);
+
+    const auto scale_after  = (_bbinfo.apply_scale() ? T(_bbinfo.scale()) : T(1));
+    const auto scale_before = T(_bbinfo.scale());
+    ARM_COMPUTE_ERROR_ON(scale_before <= 0);
+    const auto offset = (_bbinfo.correct_transform_coords() ? T(1.f) : T(0.f));
+
+    auto pred_ptr  = reinterpret_cast<T *>(_pred_boxes->buffer() + _pred_boxes->info()->offset_first_element_in_bytes());
+    auto delta_ptr = reinterpret_cast<T *>(_deltas->buffer() + _deltas->info()->offset_first_element_in_bytes());
+
+    Iterator box_it(_boxes, window);
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        const auto ptr    = reinterpret_cast<T *>(box_it.ptr());
+        const auto b0     = *ptr;
+        const auto b1     = *(ptr + 1);
+        const auto b2     = *(ptr + 2);
+        const auto b3     = *(ptr + 3);
+        const T    width  = (b2 / scale_before) - (b0 / scale_before) + T(1.f);
+        const T    height = (b3 / scale_before) - (b1 / scale_before) + T(1.f);
+        const T    ctr_x  = (b0 / scale_before) + T(0.5f) * width;
+        const T    ctr_y  = (b1 / scale_before) + T(0.5f) * height;
+        for(size_t j = 0; j < num_classes; ++j)
+        {
+            // Extract deltas
+            const size_t delta_id = id.y() * deltas_width + 4u * j;
+            const T      dx       = delta_ptr[delta_id] / T(_bbinfo.weights()[0]);
+            const T      dy       = delta_ptr[delta_id + 1] / T(_bbinfo.weights()[1]);
+            T            dw       = delta_ptr[delta_id + 2] / T(_bbinfo.weights()[2]);
+            T            dh       = delta_ptr[delta_id + 3] / T(_bbinfo.weights()[3]);
+            // Clip dw and dh
+            dw = std::min(dw, T(_bbinfo.bbox_xform_clip()));
+            dh = std::min(dh, T(_bbinfo.bbox_xform_clip()));
+            // Determine the predictions
+            const T pred_ctr_x = dx * width + ctr_x;
+            const T pred_ctr_y = dy * height + ctr_y;
+            const T pred_w     = T(std::exp(dw)) * width;
+            const T pred_h     = T(std::exp(dh)) * height;
+            // Store the prediction into the output tensor
+            pred_ptr[delta_id]     = scale_after * utility::clamp<T>(pred_ctr_x - T(0.5f) * pred_w, T(0), T(img_w - 1));
+            pred_ptr[delta_id + 1] = scale_after * utility::clamp<T>(pred_ctr_y - T(0.5f) * pred_h, T(0), T(img_h - 1));
+            pred_ptr[delta_id + 2] = scale_after * utility::clamp<T>(pred_ctr_x + T(0.5f) * pred_w - offset, T(0), T(img_w - 1));
+            pred_ptr[delta_id + 3] = scale_after * utility::clamp<T>(pred_ctr_y + T(0.5f) * pred_h - offset, T(0), T(img_h - 1));
+        }
+    },
+    box_it);
+}
+
+void NEBoundingBoxTransformKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+    switch(_boxes->info()->data_type())
+    {
+        case DataType::F32:
+        {
+            internal_run<float>(window, info);
+            break;
+        }
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        case DataType::F16:
+        {
+            internal_run<float16_t>(window, info);
+            break;
+        }
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        default:
+        {
+            ARM_COMPUTE_ERROR("Data type not supported");
+        }
+    }
+}
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEBoundingBoxTransform.cpp b/src/runtime/NEON/functions/NEBoundingBoxTransform.cpp
new file mode 100644
index 0000000000..818e228d52
--- /dev/null
+++ b/src/runtime/NEON/functions/NEBoundingBoxTransform.cpp
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/runtime/NEON/functions/NEBoundingBoxTransform.h"
+
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+void NEBoundingBoxTransform::configure(const ITensor *boxes, ITensor *pred_boxes, const ITensor *deltas, const BoundingBoxTransformInfo &info)
+{
+    // Configure Bounding Box kernel
+    auto k = arm_compute::support::cpp14::make_unique<NEBoundingBoxTransformKernel>();
+    k->configure(boxes, pred_boxes, deltas, info);
+    _kernel = std::move(k);
+}
+
+Status NEBoundingBoxTransform::validate(const ITensorInfo *boxes, const ITensorInfo *pred_boxes, const ITensorInfo *deltas, const BoundingBoxTransformInfo &info)
+{
+    return NEBoundingBoxTransformKernel::validate(boxes, pred_boxes, deltas, info);
+}
+} // namespace arm_compute