COMPMID-1794: Add support for NHWC in CLROIAlignLayer
Change-Id: If1df8f6c0549c986e607cbceb0977c80b2891b75
Reviewed-on: https://review.mlplatform.org/493
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Isabella Gottardi <isabella.gottardi@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
diff --git a/src/core/CL/cl_kernels/roi_align_layer.cl b/src/core/CL/cl_kernels/roi_align_layer.cl
index f52eb18..a956860 100644
--- a/src/core/CL/cl_kernels/roi_align_layer.cl
+++ b/src/core/CL/cl_kernels/roi_align_layer.cl
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -75,11 +75,17 @@
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);
+#if defined(NHWC)
+ const DATA_TYPE data1 = *(__global DATA_TYPE *)tensor3D_offset(input, pz, x_low, y_low);
+ const DATA_TYPE data2 = *(__global DATA_TYPE *)tensor3D_offset(input, pz, x_high, y_low);
+ const DATA_TYPE data3 = *(__global DATA_TYPE *)tensor3D_offset(input, pz, x_low, y_high);
+ const DATA_TYPE data4 = *(__global DATA_TYPE *)tensor3D_offset(input, pz, x_high, y_high);
+#else // !defined(NHWC)
+ 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);
+#endif // defined(NHWC)
sum += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4;
}
}
@@ -133,9 +139,15 @@
Image rois = CONVERT_TO_IMAGE_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);
+#if defined(NHWC)
+ const int px = get_global_id(1);
+ const int py = get_global_id(2);
+ const int pw = get_global_id(0);
+#else // !defined(NHWC)
+ const int px = get_global_id(0);
+ const int py = get_global_id(1);
+ const int pw = get_global_id(2);
+#endif // defined(NHWC)
// Load roi parameters
// roi is laid out as follows { batch_index, x1, y1, x2, y2 }
@@ -161,7 +173,7 @@
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(bin_size - EPS_GRID);
+ const float2 roi_bin_grid = ceil(bin_size - EPS_GRID);
#endif // defined(SAMPLING_RATIO)
// Move input and output pointer across the fourth dimension
@@ -169,15 +181,20 @@
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);
+#if defined(NHWC)
+ DATA_TYPE *_output_ptr = (__global DATA_TYPE *)tensor3D_offset(&output, pz, px, py);
+#else // !defined(NHWC)
+ DATA_TYPE *_output_ptr = (__global DATA_TYPE *)tensor3D_offset(&output, px, py, pz);
+#endif // defined(NHWC)
+ *_output_ptr = (__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
index 325eeb2..66d2623 100644
--- a/src/core/CL/kernels/CLROIAlignLayerKernel.cpp
+++ b/src/core/CL/kernels/CLROIAlignLayerKernel.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -34,6 +34,9 @@
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+using namespace arm_compute::misc::shape_calculator;
namespace arm_compute
{
@@ -47,18 +50,15 @@
ARM_COMPUTE_RETURN_ERROR_ON(rois->num_dimensions() > 2);
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_DATA_LAYOUT_NOT_IN(input, DataLayout::NCHW);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC, DataLayout::NCHW);
ARM_COMPUTE_RETURN_ERROR_ON((pool_info.pooled_width() == 0) || (pool_info.pooled_height() == 0));
if(output->total_size() != 0)
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(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(rois->dimension(1) != output->dimension(3));
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(compute_roi_align_shape(*input, *rois, pool_info), output->tensor_shape());
}
-
return Status{};
}
@@ -67,8 +67,9 @@
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
// Output auto inizialitation if not yet initialized
- TensorShape output_shape(pool_info.pooled_width(), pool_info.pooled_height(), input->dimension(2), rois->dimension(1));
+ const TensorShape output_shape = compute_roi_align_shape(*input, *rois, pool_info);
auto_init_if_empty((*output), output_shape, 1, input->data_type());
+ output->set_data_layout(input->data_layout());
// Configure kernel window
const unsigned int num_elems_processed_per_iteration = 1;
@@ -107,12 +108,13 @@
CLBuildOptions build_opts;
build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
build_opts.add_option("-DDATA_SIZE=" + get_data_size_from_data_type(input->info()->data_type()));
- build_opts.add_option("-DMAX_DIM_X=" + support::cpp11::to_string(_input->info()->dimension(Window::DimX)));
- build_opts.add_option("-DMAX_DIM_Y=" + support::cpp11::to_string(_input->info()->dimension(Window::DimY)));
- build_opts.add_option("-DMAX_DIM_Z=" + support::cpp11::to_string(_input->info()->dimension(Window::DimZ)));
+ build_opts.add_option("-DMAX_DIM_X=" + support::cpp11::to_string(_input->info()->dimension(get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH))));
+ build_opts.add_option("-DMAX_DIM_Y=" + support::cpp11::to_string(_input->info()->dimension(get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT))));
+ build_opts.add_option("-DMAX_DIM_Z=" + support::cpp11::to_string(_input->info()->dimension(get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::CHANNEL))));
build_opts.add_option("-DPOOLED_DIM_X=" + support::cpp11::to_string(pool_info.pooled_width()));
build_opts.add_option("-DPOOLED_DIM_Y=" + support::cpp11::to_string(pool_info.pooled_height()));
build_opts.add_option("-DSPATIAL_SCALE=" + float_to_string_with_full_precision(pool_info.spatial_scale()));
+ build_opts.add_option_if(input->info()->data_layout() == DataLayout::NHWC, "-DNHWC");
build_opts.add_option_if(pool_info.sampling_ratio() > 0, "-DSAMPLING_RATIO=" + support::cpp11::to_string(pool_info.sampling_ratio()));
// Create kernel
@@ -137,7 +139,7 @@
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]);
+ slice.set(get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::CHANNEL), window[3]);
// Set arguments
unsigned int idx = 0;