[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>

4 files changed, 374 insertions, 0 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 87b588e16c..ce4b85551d 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -358,6 +358,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "RGBA8888_to_NV12_bt709", "color_convert.cl" },
     { "RGBA8888_to_RGB888_bt709", "color_convert.cl" },
     { "RGBA8888_to_YUV444_bt709", "color_convert.cl" },
+    { "roi_align_layer", "roi_align_layer.cl" },
     { "roi_pooling_layer", "roi_pooling_layer.cl" },
     { "scale_nearest_neighbour_nchw", "scale.cl" },
     { "scale_nearest_neighbour_nhwc", "scale.cl" },
@@ -762,6 +763,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/reshape_layer.clembed"
     },
     {
+        "roi_align_layer.cl",
+#include "./cl_kernels/roi_align_layer.clembed"
+    },
+    {
         "roi_pooling_layer.cl",
 #include "./cl_kernels/roi_pooling_layer.clembed"
     },
diff --git a/src/core/CL/cl_kernels/roi_align_layer.cl b/src/core/CL/cl_kernels/roi_align_layer.cl
new file mode 100644
index 0000000000..4625e53ed5
--- /dev/null
+++ b/src/core/CL/cl_kernels/roi_align_layer.cl
@@ -0,0 +1,181 @@
+/*
+ * 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 "helpers.h"
+
+// This specifies the value to shift the result of roi_dims / pooled_dims before ceiling.
+// It is close to the epsilon machine (for a floating point system, x and x+EPS are the same number).
+#define EPS_GRID 0.00001f
+
+#if defined(DATA_TYPE) && defined(POOLED_DIM_X) && defined(POOLED_DIM_Y) && defined(MAX_DIM_X) && defined(MAX_DIM_Y) && defined(MAX_DIM_Z) && defined(SPATIAL_SCALE) // Check for compile time constants
+
+/** Performs a roi align on a single output pixel.
+ *
+ * @param[in] input          Pointer to input Tensor3D struct.
+ * @param[in] region_start_x Start x index projected onto the input tensor.
+ * @param[in] region_end_x   End x index projected onto the input tensor.
+ * @param[in] region_start_y Start y index projected onto the input tensor.
+ * @param[in] region_end_y   End y index projected onto the input tensor.
+ * @param[in] pz             z index of the input tensor.
+ *
+ * @return An average pooled value from the region specified in the input tensor.
+ */
+inline DATA_TYPE roi_align_1x1(const Tensor3D *input, float region_start_x,
+                               float bin_size_x,
+                               float grid_size_x,
+                               float region_end_x,
+                               float region_start_y,
+                               float bin_size_y,
+                               float grid_size_y,
+                               float region_end_y,
+                               int   pz)
+{
+    // Iterate through the pooling region
+    float sum = 0;
+    for(int iy = 0; iy < grid_size_y; ++iy)
+    {
+        for(int ix = 0; ix < grid_size_x; ++ix)
+        {
+            // Align the window in the middle of every bin
+            const float y = region_start_y + (iy + 0.5f) * bin_size_y / (float)grid_size_y;
+            const float x = region_start_x + (ix + 0.5f) * bin_size_x / (float)grid_size_x;
+
+            // Interpolation in the unit square
+            const int y_low  = (int)y;
+            const int x_low  = (int)x;
+            const int y_high = y_low + 1;
+            const int x_high = x_low + 1;
+
+            const float ly = y - y_low;
+            const float lx = x - x_low;
+            const float hy = 1.f - ly;
+            const float hx = 1.f - lx;
+
+            const float w1 = hy * hx;
+            const float w2 = hy * lx;
+            const float w3 = ly * hx;
+            const float w4 = ly * lx;
+
+            const DATA_TYPE data1 = *(__global DATA_TYPE *)tensor3D_offset(input, x_low, y_low, pz);
+            const DATA_TYPE data2 = *(__global DATA_TYPE *)tensor3D_offset(input, x_high, y_low, pz);
+            const DATA_TYPE data3 = *(__global DATA_TYPE *)tensor3D_offset(input, x_low, y_high, pz);
+            const DATA_TYPE data4 = *(__global DATA_TYPE *)tensor3D_offset(input, x_high, y_high, pz);
+            sum += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4;
+        }
+    }
+
+    return (DATA_TYPE)(sum / (grid_size_x * grid_size_y));
+}
+
+/** Performs a roi align function.
+ *
+ * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16, F32;
+ * @note Datasize must be passed using -DDATA_SIZE e.g. -DDATA_SIZE=32;
+ * @note Input dimensions must be passed using -DMAX_DIM_X, -DMAX_DIM_Y and -DMAX_DIM_Z;
+ * @note Pooled region dimensions must be passed using -DPOOLED_DIM_X and -DPOOLED_DIM_Y;
+ * @note Spatial scale must be passed using -DSPATIAL_SCALE;
+ * @note Sampling ratio (i.e., the number of samples in each bin) may be passed using -DSAMPLING_RATIO. If not defined each roi
+ *       will have a default sampling ratio of roi_dims/pooling_dims
+ *
+ * @param[in]  input_ptr                            Pointer to the source image. Supported data types: F16, F32
+ * @param[in]  input_stride_x                       Stride of the source image in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the source image in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the pooled region of the source image as specifed by ROI
+ * @param[in]  rois_ptr                             Pointer to the rois array. Layout: {x, y, width, height, batch_indx}
+ * @param[in]  rois_stride_x                        Stride of the rois array in X dimension (in bytes)
+ * @param[in]  rois_step_x                          rois_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  rois_offset_first_element_in_bytes   The offset of the first element in the rois array
+ * @param[out] output_ptr                           Pointer to the destination image. Supported data types: F16, F32
+ * @param[in]  output_stride_x                      Stride of the destination image in X dimension (in bytes)
+ * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[in]  input_stride_w                       Stride of the source image in W dimension (in bytes)
+ * @param[in]  output_stride_w                      Stride of the destination image in W dimension (in bytes)
+ */
+__kernel void roi_align_layer(
+    TENSOR3D_DECLARATION(input),
+    VECTOR_DECLARATION(rois),
+    TENSOR3D_DECLARATION(output),
+    unsigned int input_stride_w, unsigned int output_stride_w)
+{
+    // Get pixels pointer
+    Tensor3D input  = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(input);
+    Vector   rois   = CONVERT_TO_VECTOR_STRUCT_NO_STEP(rois);
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output);
+
+    const int px = get_global_id(0);
+    const int py = get_global_id(1);
+    const int pw = get_global_id(2);
+
+    // Load roi parameters
+    // roi is laid out as follows:
+    // { x, y, width, height, batch_index }
+    const ushort4 roi       = vload4(0, (__global ushort *)vector_offset(&rois, pw));
+    const ushort roi_batch  = *((__global ushort *)vector_offset(&rois, pw) + 4);
+    const float2 roi_anchor = convert_float2(roi.s01) * convert_float(SPATIAL_SCALE);
+    const float2 roi_dims   = fmax(convert_float2(roi.s23) * convert_float(SPATIAL_SCALE), 1.f);
+
+    // Calculate pooled region start and end
+    const float2 spatial_indx     = (float2)(px, py);
+    const float2 pooled_dims      = (float2)(POOLED_DIM_X, POOLED_DIM_Y);
+    const float2 max_spatial_dims = (float2)(MAX_DIM_X, MAX_DIM_Y);
+
+    const float2 bin_size     = roi_dims / pooled_dims;
+    float2       region_start = spatial_indx * bin_size + roi_anchor;
+    float2       region_end   = (spatial_indx + 1) * bin_size + roi_anchor;
+
+    region_start = clamp(region_start, 0, max_spatial_dims);
+    region_end   = clamp(region_end, 0, max_spatial_dims);
+
+#if defined(SAMPLING_RATIO)
+    const float2 roi_bin_grid = SAMPLING_RATIO;
+#else  // !defined(SAMPLING_RATIO)
+    // Note that we subtract EPS_GRID before ceiling. This is to avoid situations where 1.000001 gets ceiled to 2.
+    const float2 roi_bin_grid = ceil(roi_dims / pooled_dims - EPS_GRID);
+#endif // defined(SAMPLING_RATIO)
+
+    // Move input and output pointer across the fourth dimension
+    input.ptr += roi_batch * input_stride_w;
+    output.ptr += pw * output_stride_w;
+    for(int pz = 0; pz < MAX_DIM_Z; ++pz)
+    {
+        *(__global DATA_TYPE *)tensor3D_offset(&output, px, py, pz) = (__global DATA_TYPE)roi_align_1x1(&input,
+                                                                                                        region_start.x,
+                                                                                                        bin_size.x,
+                                                                                                        roi_bin_grid.x,
+                                                                                                        region_end.x,
+                                                                                                        region_start.y,
+                                                                                                        bin_size.y,
+                                                                                                        roi_bin_grid.y,
+                                                                                                        region_end.y, pz);
+    }
+}
+#endif // Check for compile time constants
diff --git a/src/core/CL/kernels/CLROIAlignLayerKernel.cpp b/src/core/CL/kernels/CLROIAlignLayerKernel.cpp
new file mode 100644
index 0000000000..2e1e85488b
--- /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
diff --git a/src/runtime/CL/functions/CLROIAlignLayer.cpp b/src/runtime/CL/functions/CLROIAlignLayer.cpp
new file mode 100644
index 0000000000..1528759840
--- /dev/null
+++ b/src/runtime/CL/functions/CLROIAlignLayer.cpp
@@ -0,0 +1,47 @@
+/*
+ * 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/runtime/CL/functions/CLROIAlignLayer.h"
+
+#include "arm_compute/core/CL/ICLArray.h"
+#include "arm_compute/core/CL/kernels/CLROIAlignLayerKernel.h"
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+Status CLROIAlignLayer::validate(const ITensorInfo *input, size_t num_rois, ITensorInfo *output, const ROIPoolingLayerInfo &pool_info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(CLROIAlignLayerKernel::validate(input, num_rois, output, pool_info));
+
+    return Status{};
+}
+
+void CLROIAlignLayer::configure(const ICLTensor *input, const ICLROIArray *rois, ICLTensor *output, const ROIPoolingLayerInfo &pool_info)
+{
+    // Configure ROI pooling kernel
+    auto k = arm_compute::support::cpp14::make_unique<CLROIAlignLayerKernel>();
+    k->configure(input, rois, output, pool_info);
+    _kernel = std::move(k);
+}
+
+} // namespace arm_compute