COMPMID-1527 - Implementing ReorgLayer on OpenCL

Also extended tests on NEON

Change-Id: Icb0eced534e904ef807972dd3a31988f501bb02e
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/147095
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

5 files changed, 272 insertions, 6 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 4af2b09530..7cc586bff1 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -336,6 +336,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "reduction_operation", "reduction_operation.cl" },
     { "remap_nearest_neighbour", "remap.cl" },
     { "remap_bilinear", "remap.cl" },
+    { "reorg_layer_nchw", "reorg_layer.cl" },
+    { "reorg_layer_nhwc", "reorg_layer.cl" },
     { "reshape_layer", "reshape_layer.cl" },
     { "reshape_to_columns", "convolution_layer.cl" },
     { "RGB888_to_IYUV_bt709", "color_convert.cl" },
@@ -720,6 +722,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/remap.clembed"
     },
     {
+        "reorg_layer.cl",
+#include "./cl_kernels/reorg_layer.clembed"
+    },
+    {
         "reshape_layer.cl",
 #include "./cl_kernels/reshape_layer.clembed"
     },
diff --git a/src/core/CL/cl_kernels/reorg_layer.cl b/src/core/CL/cl_kernels/reorg_layer.cl
new file mode 100644
index 0000000000..a275699a3f
--- /dev/null
+++ b/src/core/CL/cl_kernels/reorg_layer.cl
@@ -0,0 +1,116 @@
+/*
+ * 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"
+
+#if defined(DATA_TYPE) && defined(SRC_DEPTH) && defined(STRIDE)
+
+#define CALCULATE_SRC_COORDINATES(xo, yo, zo, xi, yi, zi)     \
+    ({                                                        \
+        int offset = zo / (int)SRC_DEPTH;                     \
+        xi         = xo * (int)STRIDE + offset % (int)STRIDE; \
+        yi         = yo * (int)STRIDE + offset / (int)STRIDE; \
+        zi         = zo % SRC_DEPTH;                          \
+    })
+
+/** Performs a reorganization layer of input tensor to the output tensor when the data layout is NCHW
+ *
+ * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
+ * @note The depth of the input tensor must be passed at compile time using -DSRC_DEPTH: e.g. -DSRC_DEPTH=64
+ * @note The distance between 2 consecutive pixels along the x and y direction must be passed at compile time using -DSTRIDE: e.g. -DSTRIDE=2
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void reorg_layer_nchw(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    int xo = get_global_id(0);
+    int yo = get_global_id(1);
+    int zo = get_global_id(2);
+    int xi, yi, zi;
+
+    CALCULATE_SRC_COORDINATES(xo, yo, zo, xi, yi, zi);
+
+    int src_offset                   = xi * sizeof(DATA_TYPE) + yi * src_stride_y + zi * src_stride_z;
+    *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)(src_ptr + src_offset_first_element_in_bytes + src_offset));
+}
+
+/** Performs a reorganization layer of input tensor to the output tensor when the data layout is NHWC
+ *
+ * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
+ * @note The depth of the input tensor must be passed at compile time using -DSRC_DEPTH: e.g. -DSRC_DEPTH=64
+ * @note The distance between 2 consecutive pixels along the x and y direction must be passed at compile time using -DSTRIDE: e.g. -DSTRIDE=2
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void reorg_layer_nhwc(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
+{
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    int xo = get_global_id(1);
+    int yo = get_global_id(2);
+    int zo = get_global_id(0);
+    int xi, yi, zi;
+
+    CALCULATE_SRC_COORDINATES(xo, yo, zo, xi, yi, zi);
+
+    int src_offset = zi * sizeof(DATA_TYPE) + xi * src_stride_y + yi * src_stride_z;
+
+    *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)(src_ptr + src_offset_first_element_in_bytes + src_offset));
+}
+#endif // // defined(DATA_TYPE) && defined(SRC_DEPTH) && defined(STRIDE)
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp
index d4a72076c1..1fa5c8521f 100644
--- a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp
@@ -192,8 +192,6 @@ void CLBatchNormalizationLayerKernel::configure(ICLTensor *input, ICLTensor *out
     ICLKernel::configure_internal(win_config.second);
 
     _config_id = "batch_normalization_layer_";
-    _config_id += string_from_data_layout(input->info()->data_layout());
-    _config_id += "_";
     _config_id += string_from_data_type(input->info()->data_type());
     _config_id += "_";
     _config_id += support::cpp11::to_string(input->info()->dimension(0));
diff --git a/src/core/CL/kernels/CLReorgLayerKernel.cpp b/src/core/CL/kernels/CLReorgLayerKernel.cpp
new file mode 100644
index 0000000000..7891844ef6
--- /dev/null
+++ b/src/core/CL/kernels/CLReorgLayerKernel.cpp
@@ -0,0 +1,140 @@
+/*
+ * 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/CLReorgLayerKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include <string>
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, int32_t stride)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1,
+                                                         DataType::U8, DataType::S8, DataType::QASYMM8,
+                                                         DataType::U16, DataType::S16,
+                                                         DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
+
+    const size_t idx_width  = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
+
+    ARM_COMPUTE_RETURN_ERROR_ON(stride <= 0);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((input->tensor_shape()[idx_width] % stride) != 0, "The width of the input tensor must be a multiple of stride");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((input->tensor_shape()[idx_height] % stride) != 0, "The height of the input tensor must be a multiple of stride");
+
+    // Validate output if initialized
+    if(output->total_size() != 0)
+    {
+        const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(misc::shape_calculator::compute_reorg_output_shape(*input, stride));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+} // namespace
+
+CLReorgLayerKernel::CLReorgLayerKernel()
+    : _input(nullptr), _output(nullptr)
+{
+}
+
+void CLReorgLayerKernel::configure(const ICLTensor *input, ICLTensor *output, int32_t stride)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), stride));
+
+    _input  = input;
+    _output = output;
+
+    std::string  kernel_name = std::string("reorg_layer_") + lower_string(string_from_data_layout(input->info()->data_layout()));
+    const size_t idx_channel = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::CHANNEL);
+
+    // Create kernel
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(input->info()->dimension(idx_channel)));
+    build_opts.add_option("-DSTRIDE=" + support::cpp11::to_string(stride));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+
+    // Configure window
+    // auto inizialize the output tensor if not yet initialized
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(misc::shape_calculator::compute_reorg_output_shape(*input->info(), stride)));
+
+    Window win = calculate_max_window(*output->info(), Steps());
+
+    // The CLWeightsReshapeKernel doesn't need padding so update_window_and_padding() can be skipped
+    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
+    ICLKernel::configure_internal(win);
+
+    _config_id = kernel_name;
+    _config_id += "_";
+    _config_id += string_from_data_type(input->info()->data_type());
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(1));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(2));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(stride);
+}
+
+Status CLReorgLayerKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output, int32_t stride)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, stride));
+
+    return Status{};
+}
+
+void CLReorgLayerKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window slice = window.first_slice_window_3D();
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, _output, slice);
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    while(window.slide_window_slice_3D(slice));
+}
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/NEON/kernels/NEReorgLayerKernel.cpp b/src/core/NEON/kernels/NEReorgLayerKernel.cpp
index 1b2ec92cba..8baea2b990 100644
--- a/src/core/NEON/kernels/NEReorgLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEReorgLayerKernel.cpp
@@ -46,14 +46,20 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, i
                                                          DataType::U16, DataType::S16,
                                                          DataType::U32, DataType::S32,
                                                          DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(stride <= 0, "Stride should be a positive number");
+    ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
 
-    const TensorShape output_shape = misc::shape_calculator::compute_reorg_output_shape(*input, stride);
+    const size_t idx_width  = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
+    const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
 
-    // Validate configured output
+    ARM_COMPUTE_RETURN_ERROR_ON(stride <= 0);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((input->tensor_shape()[idx_width] % stride) != 0, "The width of the input tensor must be a multiple of stride");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((input->tensor_shape()[idx_height] % stride) != 0, "The height of the input tensor must be a multiple of stride");
+
+    // Validate output if initialized
     if(output->total_size() != 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
+        const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(misc::shape_calculator::compute_reorg_output_shape(*input, stride));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     }