COMPMID-1677: Change ROIPooling layer interface to accept ROIs as tensors

Change-Id: If16b572a4d906187b77f32133a72a44316fa74e4
Reviewed-on: https://review.mlplatform.org/490
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 54 insertions, 28 deletions

diff --git a/src/core/CL/kernels/CLROIPoolingLayerKernel.cpp b/src/core/CL/kernels/CLROIPoolingLayerKernel.cpp
index 23676942a6..df7687edea 100644
--- a/src/core/CL/kernels/CLROIPoolingLayerKernel.cpp
+++ b/src/core/CL/kernels/CLROIPoolingLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -39,29 +39,61 @@
 #include <set>
 #include <string>
 
-using namespace arm_compute;
+namespace arm_compute
+{
+namespace
+{
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *rois, ITensorInfo *output, const ROIPoolingLayerInfo &pool_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    // Output auto initialization if not yet initialized
+    TensorShape output_shape(pool_info.pooled_width(), pool_info.pooled_height(), input->dimension(2), rois->dimension(1));
+    auto_init_if_empty((*output), output_shape, 1, input->data_type());
+
+    // Configure kernel window
+    const unsigned int num_elems_processed_per_iteration = 1;
+    Window             win                               = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
+
+    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
+    AccessWindowHorizontal input_access(input, input->valid_region().start(0), num_elems_processed_per_iteration);
+
+    bool window_changed = update_window_and_padding(win, input_access, output_access);
+    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
 
 CLROIPoolingLayerKernel::CLROIPoolingLayerKernel()
     : _input(nullptr), _rois(nullptr), _output(nullptr), _pool_info(0, 0, 0.f)
 {
 }
 
-void CLROIPoolingLayerKernel::configure(const ICLTensor *input, const ICLROIArray *rois, ICLTensor *output, const ROIPoolingLayerInfo &pool_info)
+void CLROIPoolingLayerKernel::configure(const ICLTensor *input, const ICLTensor *rois, ICLTensor *output, const ROIPoolingLayerInfo &pool_info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, rois, output);
+
+    //Validate arguments
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input->info(), rois->info(), output->info());
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(rois, 1, DataType::U16);
+    ARM_COMPUTE_ERROR_ON(rois->info()->dimension(0) != 5);
+    ARM_COMPUTE_ERROR_ON(rois->info()->num_dimensions() > 2);
     ARM_COMPUTE_ERROR_ON_F16_UNSUPPORTED(input);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
     ARM_COMPUTE_ERROR_ON((pool_info.pooled_width() == 0) || (pool_info.pooled_height() == 0));
-    ARM_COMPUTE_ERROR_ON(rois->num_values() == 0);
 
-    // Output auto inizialitation if not yet initialized
-    TensorShape output_shape(pool_info.pooled_width(), pool_info.pooled_height(), input->info()->dimension(2), rois->num_values());
-    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type());
+    if(output->info()->total_size() != 0)
+    {
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) != pool_info.pooled_width()) || (output->info()->dimension(1) != pool_info.pooled_height()));
+        ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(2));
+        ARM_COMPUTE_ERROR_ON(rois->info()->dimension(1) != output->info()->dimension(3));
+    }
 
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) != pool_info.pooled_width()) || (output->info()->dimension(1) != pool_info.pooled_height()));
-    ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(2));
-    ARM_COMPUTE_ERROR_ON(rois->num_values() != output->info()->dimension(3));
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), rois->info(), output->info(), pool_info);
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
 
     // Set instance variables
     _input     = input;
@@ -89,19 +121,7 @@ void CLROIPoolingLayerKernel::configure(const ICLTensor *input, const ICLROIArra
     add_argument<cl_uint>(idx, _input->info()->strides_in_bytes()[3]);
     add_argument<cl_uint>(idx, _output->info()->strides_in_bytes()[3]);
 
-    // Configure kernel window
-    const unsigned int num_elems_processed_per_iteration = 1;
-    Window             window                            = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
-    AccessWindowStatic input_access(input->info(),
-                                    input->info()->valid_region().start(0),
-                                    input->info()->valid_region().start(1),
-                                    input->info()->valid_region().end(0),
-                                    input->info()->valid_region().end(1));
-    AccessWindowStatic output_access(output->info(), 0, 0, pool_info.pooled_width(), pool_info.pooled_height());
-
-    update_window_and_padding(window, input_access, output_access);
-    output_access.set_valid_region(window, ValidRegion(Coordinates(), output->info()->tensor_shape()));
-    ICLKernel::configure_internal(window);
+    ICLKernel::configure_internal(win_config.second);
 }
 
 void CLROIPoolingLayerKernel::run(const Window &window, cl::CommandQueue &queue)
@@ -109,14 +129,20 @@ void CLROIPoolingLayerKernel::run(const Window &window, cl::CommandQueue &queue)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
 
-    Window slice = window.first_slice_window_3D();
-    // Parallelize spatially and across the fourth dimension of the output tensor (also across ROIArray)
+    Window slice      = window.first_slice_window_3D();
+    Window slice_rois = slice;
+    // Parallelize spatially and across the fourth dimension of the output tensor (also across ROITensor)
+    slice_rois.set_dimension_step(Window::DimX, _rois->info()->dimension(0));
     slice.set(Window::DimZ, window[3]);
 
     // Set arguments
     unsigned int idx = 0;
     add_3D_tensor_argument(idx, _input, slice);
-    add_1D_array_argument<ROI>(idx, _rois, Strides(sizeof(ROI)), 1U, slice);
+    add_2D_tensor_argument(idx, _rois, slice_rois);
     add_3D_tensor_argument(idx, _output, slice);
+    add_argument<cl_uint>(idx, _input->info()->strides_in_bytes()[3]);
+    add_argument<cl_uint>(idx, _output->info()->strides_in_bytes()[3]);
+
     enqueue(queue, *this, slice);
 }
+} // namespace arm_compute
\ No newline at end of file