COMPMID-1999: Add support for GenerateProposals operator in CL

Change-Id: Ie08a6874347085f96b00f25bdb605eee7d683c25
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/719
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

5 files changed, 243 insertions, 17 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index a7d371dabc..4ecb885440 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -307,6 +307,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemmlowp_output_stage_quantize_down", "gemmlowp.cl" },
     { "gemmlowp_output_stage_quantize_down_fixedpoint", "gemmlowp.cl" },
     { "gemmlowp_output_stage_quantize_down_float", "gemmlowp.cl" },
+    { "generate_proposals_compute_all_anchors", "generate_proposals.cl" },
     { "harris_score_3x3", "harris_corners.cl" },
     { "harris_score_5x5", "harris_corners.cl" },
     { "harris_score_7x7", "harris_corners.cl" },
@@ -706,6 +707,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/gemv.clembed"
     },
     {
+        "generate_proposals.cl",
+#include "./cl_kernels/generate_proposals.clembed"
+    },
+    {
         "harris_corners.cl",
 #include "./cl_kernels/harris_corners.clembed"
     },
diff --git a/src/core/CL/cl_kernels/bounding_box_transform.cl b/src/core/CL/cl_kernels/bounding_box_transform.cl
index 77db5d9311..e6f470a962 100644
--- a/src/core/CL/cl_kernels/bounding_box_transform.cl
+++ b/src/core/CL/cl_kernels/bounding_box_transform.cl
@@ -28,11 +28,11 @@
 /** Perform a padded copy of input tensor to the output tensor. Padding values are defined at compile time
  *
  * @attention The following variables must be passed at compile time:
- * -# -DDATA_TYPE = Tensor data type. Supported data types: F16/F32
+ * -# -DDATA_TYPE= Tensor data type. Supported data types: F16/F32
  * -# -DWEIGHT{X,Y,W,H}= Weights [wx, wy, ww, wh] for the deltas
  * -# -DIMG_WIDTH= Original image width
  * -# -DIMG_HEIGHT= Original image height
- * -# -DBOX_FIELDS=Number of fields that are used to represent a box in boxes
+ * -# -DBOX_FIELDS= Number of fields that are used to represent a box in boxes
  *
  * @param[in]  boxes_ptr                                Pointer to the boxes tensor. Supported data types: F16/F32
  * @param[in]  boxes_stride_x                           Stride of the boxes tensor in X dimension (in bytes)
diff --git a/src/core/CL/cl_kernels/generate_proposals.cl b/src/core/CL/cl_kernels/generate_proposals.cl
new file mode 100644
index 0000000000..a947dad523
--- /dev/null
+++ b/src/core/CL/cl_kernels/generate_proposals.cl
@@ -0,0 +1,88 @@
+/*
+ * 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 "helpers.h"
+
+/** Generate all the region of interests based on the image size and the anchors passed in. For each element (x,y) of the
+ * grid, it will generate NUM_ANCHORS rois, given by shifting the grid position to match the anchor.
+ *
+ * @attention The following variables must be passed at compile time:
+ * -# -DDATA_TYPE= Tensor data type. Supported data types: F16/F32
+ * -# -DHEIGHT= Height of the feature map on which this kernel is applied
+ * -# -DWIDTH= Width of the feature map on which this kernel is applied
+ * -# -DNUM_ANCHORS= Number of anchors to be used to generate the rois per each pixel
+ * -# -DSTRIDE= Stride to be applied at each different pixel position (i.e., x_range = (1:WIDTH)*STRIDE and y_range = (1:HEIGHT)*STRIDE
+ * -# -DNUM_ROI_FIELDS= Number of fields used to represent a roi
+ *
+ * @param[in]  anchors_ptr                           Pointer to the anchors tensor. Supported data types: F16/F32
+ * @param[in]  anchors_stride_x                      Stride of the anchors tensor in X dimension (in bytes)
+ * @param[in]  anchors_step_x                        anchors_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  anchors_stride_y                      Stride of the anchors tensor in Y dimension (in bytes)
+ * @param[in]  anchors_step_y                        anchors_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  anchors_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  anchors_step_z                        anchors_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  anchors_offset_first_element_in_bytes The offset of the first element in the boxes tensor
+ * @param[out] rois_ptr                              Pointer to the rois. Supported data types: same as @p in_ptr
+ * @param[out] rois_stride_x                         Stride of the rois in X dimension (in bytes)
+ * @param[out] rois_step_x                           pred_boxes_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[out] rois_stride_y                         Stride of the rois in Y dimension (in bytes)
+ * @param[out] rois_step_y                           pred_boxes_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[out] rois_stride_z                         Stride of the rois in Z dimension (in bytes)
+ * @param[out] rois_step_z                           pred_boxes_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[out] rois_offset_first_element_in_bytes    The offset of the first element in the rois
+ */
+#if defined(DATA_TYPE) && defined(WIDTH) && defined(HEIGHT) && defined(NUM_ANCHORS) && defined(STRIDE) && defined(NUM_ROI_FIELDS)
+__kernel void generate_proposals_compute_all_anchors(
+    VECTOR_DECLARATION(anchors),
+    VECTOR_DECLARATION(rois))
+{
+    Vector anchors = CONVERT_TO_VECTOR_STRUCT_NO_STEP(anchors);
+    Vector rois    = CONVERT_TO_VECTOR_STRUCT(rois);
+
+    const size_t idx = get_global_id(0);
+    // Find the index of the anchor
+    const size_t anchor_idx = idx % NUM_ANCHORS;
+
+    // Find which shift is this thread using
+    const size_t shift_idx = idx / NUM_ANCHORS;
+
+    // Compute the shift on the X and Y direction (the shift depends exclusively by the index thread id)
+    const DATA_TYPE
+    shift_x = (DATA_TYPE)(shift_idx % WIDTH) * STRIDE;
+    const DATA_TYPE
+    shift_y = (DATA_TYPE)(shift_idx / WIDTH) * STRIDE;
+
+    const VEC_DATA_TYPE(DATA_TYPE, NUM_ROI_FIELDS)
+    shift = (VEC_DATA_TYPE(DATA_TYPE, NUM_ROI_FIELDS))(shift_x, shift_y, shift_x, shift_y);
+
+    // Read the given anchor
+    const VEC_DATA_TYPE(DATA_TYPE, NUM_ROI_FIELDS)
+    anchor = vload4(0, (__global DATA_TYPE *)vector_offset(&anchors, anchor_idx * NUM_ROI_FIELDS));
+
+    // Apply the shift to the anchor
+    const VEC_DATA_TYPE(DATA_TYPE, NUM_ROI_FIELDS)
+    shifted_anchor = anchor + shift;
+
+    vstore4(shifted_anchor, 0, (__global DATA_TYPE *)rois.ptr);
+}
+#endif //defined(DATA_TYPE) && defined(WIDTH) && defined(HEIGHT) && defined(NUM_ANCHORS) && defined(STRIDE) && defined(NUM_ROI_FIELDS)
diff --git a/src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp b/src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp
new file mode 100644
index 0000000000..f16422f815
--- /dev/null
+++ b/src/core/CL/kernels/CLGenerateProposalsLayerKernel.cpp
@@ -0,0 +1,128 @@
+/*
+ * 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/CL/kernels/CLGenerateProposalsLayerKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLArray.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/OpenCL.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"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *anchors, const ITensorInfo *all_anchors, const ComputeAnchorsInfo &info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(anchors, all_anchors);
+    ARM_COMPUTE_RETURN_ERROR_ON(anchors->dimension(0) != info.values_per_roi());
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(anchors, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(anchors->num_dimensions() > 2);
+    if(all_anchors->total_size() > 0)
+    {
+        size_t feature_height = info.feat_height();
+        size_t feature_width  = info.feat_width();
+        size_t num_anchors    = anchors->dimension(1);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(all_anchors, anchors);
+        ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->num_dimensions() > 2);
+        ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(0) != info.values_per_roi());
+        ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(1) != feature_height * feature_width * num_anchors);
+    }
+    return Status{};
+}
+} // namespace
+
+CLComputeAllAnchorsKernel::CLComputeAllAnchorsKernel()
+    : _anchors(nullptr), _all_anchors(nullptr)
+{
+}
+
+void CLComputeAllAnchorsKernel::configure(const ICLTensor *anchors, ICLTensor *all_anchors, const ComputeAnchorsInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(anchors, all_anchors);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(anchors->info(), all_anchors->info(), info));
+
+    // Metadata
+    const size_t   num_anchors = anchors->info()->dimension(1);
+    const DataType data_type   = anchors->info()->data_type();
+    const float    width       = info.feat_width();
+    const float    height      = info.feat_height();
+
+    // Initialize the output if empty
+    const TensorShape output_shape(info.values_per_roi(), width * height * num_anchors);
+    auto_init_if_empty(*all_anchors->info(), output_shape, 1, data_type);
+
+    // Set instance variables
+    _anchors     = anchors;
+    _all_anchors = all_anchors;
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
+    build_opts.add_option("-DWIDTH=" + float_to_string_with_full_precision(width));
+    build_opts.add_option("-DHEIGHT=" + float_to_string_with_full_precision(height));
+    build_opts.add_option("-DSTRIDE=" + float_to_string_with_full_precision(1.f / info.spatial_scale()));
+    build_opts.add_option("-DNUM_ANCHORS=" + support::cpp11::to_string(num_anchors));
+    build_opts.add_option("-DNUM_ROI_FIELDS=" + support::cpp11::to_string(info.values_per_roi()));
+
+    // Create kernel
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("generate_proposals_compute_all_anchors", build_opts.options()));
+
+    // The tensor all_anchors can be interpreted as an array of structs (each structs has values_per_roi fields).
+    // This means we don't need to pad on the X dimension, as we know in advance how many fields
+    // compose the struct.
+    Window win = calculate_max_window(*all_anchors->info(), Steps(info.values_per_roi()));
+    ICLKernel::configure_internal(win);
+}
+
+Status CLComputeAllAnchorsKernel::validate(const ITensorInfo *anchors, const ITensorInfo *all_anchors, const ComputeAnchorsInfo &info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(anchors, all_anchors, info));
+    return Status{};
+}
+
+void CLComputeAllAnchorsKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
+
+    // Collapse everything on the first dimension
+    Window collapsed = window.collapse(ICLKernel::window(), Window::DimX);
+
+    // Set arguments
+    unsigned int idx = 0;
+    add_1D_tensor_argument(idx, _anchors, collapsed);
+    add_1D_tensor_argument(idx, _all_anchors, collapsed);
+
+    // Note that we don't need to loop over the slices, as we are launching exactly
+    // as many threads as all the anchors generated
+    enqueue(queue, *this, collapsed);
+}
+} // namespace arm_compute
diff --git a/src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp b/src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp
index 5e4b80aa5a..02150ff275 100644
--- a/src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp
+++ b/src/core/CPP/kernels/CPPBoxWithNonMaximaSuppressionLimitKernel.cpp
@@ -54,7 +54,7 @@ std::vector<int> SoftNMS(const ITensor *proposals, std::vector<std::vector<T>> &
         areas[i] = (x2[i] - x1[i] + 1.0) * (y2[i] - y1[i] + 1.0);
     }
 
-    // Note: Soft NMS scores have already been initialize with input scores
+    // Note: Soft NMS scores have already been initialized with input scores
 
     while(!inds.empty())
     {
@@ -150,17 +150,21 @@ std::vector<int> NonMaximaSuppression(const ITensor *proposals, std::vector<int>
 
         for(unsigned int j = 0; j < sorted_indices_temp.size(); ++j)
         {
-            const auto xx1 = std::max(x1[sorted_indices_temp.at(j)], x1[i]);
-            const auto yy1 = std::max(y1[sorted_indices_temp.at(j)], y1[i]);
-            const auto xx2 = std::min(x2[sorted_indices_temp.at(j)], x2[i]);
-            const auto yy2 = std::min(y2[sorted_indices_temp.at(j)], y2[i]);
-
-            const auto w     = std::max((xx2 - xx1 + 1.f), 0.f);
-            const auto h     = std::max((yy2 - yy1 + 1.f), 0.f);
-            const auto inter = w * h;
-            const auto ovr   = inter / (areas[i] + areas[sorted_indices_temp.at(j)] - inter);
-
-            if(ovr <= info.nms())
+            const float xx1 = std::max(x1[sorted_indices_temp.at(j)], x1[i]);
+            const float yy1 = std::max(y1[sorted_indices_temp.at(j)], y1[i]);
+            const float xx2 = std::min(x2[sorted_indices_temp.at(j)], x2[i]);
+            const float yy2 = std::min(y2[sorted_indices_temp.at(j)], y2[i]);
+
+            const float w     = std::max((xx2 - xx1 + 1.f), 0.f);
+            const float h     = std::max((yy2 - yy1 + 1.f), 0.f);
+            const float inter = w * h;
+            const float ovr   = inter / (areas[i] + areas[sorted_indices_temp.at(j)] - inter);
+            const float ctr_x = xx1 + (w / 2);
+            const float ctr_y = yy1 + (h / 2);
+
+            // If suppress_size is specified, filter the boxes based on their size and position
+            const bool keep_size = !info.suppress_size() || (w >= info.min_size() && h >= info.min_size() && ctr_x < info.im_width() && ctr_y < info.im_height());
+            if(ovr <= info.nms() && keep_size)
             {
                 new_indices.push_back(j);
             }
@@ -214,8 +218,9 @@ void CPPBoxWithNonMaximaSuppressionLimitKernel::run_nmslimit()
     for(int b = 0; b < batch_size; ++b)
     {
         const int num_boxes = _batch_splits_in == nullptr ? 1 : static_cast<int>(*reinterpret_cast<T *>(_batch_splits_in->ptr_to_element(Coordinates(b))));
-        // Skip first class
-        for(int j = 1; j < num_classes; ++j)
+        // Skip first class if there is more than 1 except if the number of classes is 1.
+        const int j_start = (num_classes == 1 ? 0 : 1);
+        for(int j = j_start; j < num_classes; ++j)
         {
             std::vector<T>   cur_scores(scores_count);
             std::vector<int> inds;
@@ -290,7 +295,7 @@ void CPPBoxWithNonMaximaSuppressionLimitKernel::run_nmslimit()
 
         // Write results
         int cur_out_idx = 0;
-        for(int j = 1; j < num_classes; ++j)
+        for(int j = j_start; j < num_classes; ++j)
         {
             auto     &cur_keep        = keeps[j];
             auto      cur_out_scores  = reinterpret_cast<T *>(_scores_out->ptr_to_element(Coordinates(cur_start_idx + cur_out_idx)));