[COMPMID-1331] Add support for RoIAlign operator in CL
Change-Id: Ie215daacd10477309dbf8af1bb2b05b7a0a8f203
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/150773
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Isabella Gottardi <isabella.gottardi@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
diff --git a/src/core/CL/kernels/CLROIAlignLayerKernel.cpp b/src/core/CL/kernels/CLROIAlignLayerKernel.cpp
new file mode 100644
index 0000000..2e1e854
--- /dev/null
+++ b/src/core/CL/kernels/CLROIAlignLayerKernel.cpp
@@ -0,0 +1,141 @@
+/*
+ * Copyright (c) 2018 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/CLROIAlignLayerKernel.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 *input, size_t num_rois, const ITensorInfo *output, const ROIPoolingLayerInfo &pool_info)
+{
+ ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+ ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
+ ARM_COMPUTE_RETURN_ERROR_ON((pool_info.pooled_width() == 0) || (pool_info.pooled_height() == 0));
+ ARM_COMPUTE_RETURN_ERROR_ON(num_rois == 0);
+
+ if(output->total_size() != 0)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+ ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(0) != pool_info.pooled_width()) || (output->dimension(1) != pool_info.pooled_height()));
+ ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) != output->dimension(2));
+ ARM_COMPUTE_RETURN_ERROR_ON(num_rois != output->dimension(3));
+ }
+
+ return Status{};
+}
+} // namespace
+
+CLROIAlignLayerKernel::CLROIAlignLayerKernel()
+ : _input(nullptr), _output(nullptr), _rois(nullptr), _pool_info(0, 0, 0.f)
+{
+}
+
+void CLROIAlignLayerKernel::configure(const ICLTensor *input, const ICLROIArray *rois, ICLTensor *output, const ROIPoolingLayerInfo &pool_info)
+{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, rois);
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), rois->num_values(), output->info(), pool_info));
+
+ 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());
+
+ _input = input;
+ _output = output;
+ _rois = rois;
+ _pool_info = pool_info;
+
+ // Set build options
+ std::set<std::string> build_opts;
+ build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())));
+ build_opts.emplace(("-DDATA_SIZE=" + get_data_size_from_data_type(input->info()->data_type())));
+ build_opts.emplace(("-DMAX_DIM_X=" + support::cpp11::to_string(_input->info()->dimension(Window::DimX))));
+ build_opts.emplace(("-DMAX_DIM_Y=" + support::cpp11::to_string(_input->info()->dimension(Window::DimY))));
+ build_opts.emplace(("-DMAX_DIM_Z=" + support::cpp11::to_string(_input->info()->dimension(Window::DimZ))));
+ build_opts.emplace(("-DPOOLED_DIM_X=" + support::cpp11::to_string(pool_info.pooled_width())));
+ build_opts.emplace(("-DPOOLED_DIM_Y=" + support::cpp11::to_string(pool_info.pooled_height())));
+ build_opts.emplace(("-DSPATIAL_SCALE=" + float_to_string_with_full_precision(pool_info.spatial_scale())));
+ if(pool_info.sampling_ratio() > 0)
+ {
+ build_opts.emplace(("-DSAMPLING_RATIO=" + support::cpp11::to_string(pool_info.sampling_ratio())));
+ }
+
+ // Create kernel
+ std::string kernel_name = "roi_align_layer";
+ _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts));
+
+ // Set static kernel arguments
+ unsigned int idx = 2 * num_arguments_per_3D_tensor() + num_arguments_per_1D_array();
+ 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());
+
+ output_access.set_valid_region(window, ValidRegion(Coordinates(), output->info()->tensor_shape()));
+ ICLKernel::configure_internal(window);
+}
+
+Status CLROIAlignLayerKernel::validate(const ITensorInfo *input, size_t num_rois, ITensorInfo *output, const ROIPoolingLayerInfo &pool_info)
+{
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, num_rois, output, pool_info));
+ return Status{};
+}
+
+void CLROIAlignLayerKernel::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)
+ 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_3D_tensor_argument(idx, _output, slice);
+ enqueue(queue, *this, slice);
+}
+
+} // namespace arm_compute