COMPMID-2239 Implement DEPTH_TO_SPACE for CL

Change-Id: Ia064b742624c0052a7e2ff9bebe84e675369ff86
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1233
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 245 insertions, 0 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 23eac1664a..c0031cb546 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -225,6 +225,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "dwc_3x3_reshaped_qasymm8_nhwc", "depthwise_convolution_quantized.cl" },
     { "dwc_3x3_reshaped_qasymm8_stride1_nhwc", "depthwise_convolution_quantized.cl" },
     { "dwc_3x3_reshaped_qasymm8_dot8_stride1_nhwc", "depthwise_convolution_quantized.cl" },
+    { "depth_to_space_nchw", "depth_to_space.cl" },
+    { "depth_to_space_nhwc", "depth_to_space.cl" },
     { "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16", "depthwise_convolution.cl" },
     { "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16", "depthwise_convolution.cl" },
     { "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32", "depthwise_convolution.cl" },
@@ -668,6 +670,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/depth_convert.clembed"
     },
     {
+        "depth_to_space.cl",
+#include "./cl_kernels/depth_to_space.clembed"
+    },
+    {
         "depthwise_convolution.cl",
 #include "./cl_kernels/depthwise_convolution.clembed"
     },
diff --git a/src/core/CL/cl_kernels/depth_to_space.cl b/src/core/CL/cl_kernels/depth_to_space.cl
new file mode 100644
index 0000000000..2ffd0a40e7
--- /dev/null
+++ b/src/core/CL/cl_kernels/depth_to_space.cl
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2019 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(BLOCK_SHAPE) && defined(CHANNEL_SIZE)
+/** Batch to space transformation. (NCHW)
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note The input tensor batch size must be passed at compile time using -DCHANNEL_SIZE. e.g. -DCHANNEL_SIZE=2
+ * @note The block shape must be passed at compile time using -DBLOCK_SHAPE. e.g. -DBLOCK_SHAPE=2
+ *
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  input_stride_x                       Stride of the source tensor 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 tensor 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 first source tensor
+ * @param[in]  batch_id                             The input tensor batch id
+ * @param[out] output_ptr                           Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination tensor 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 tensor 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 source 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 tensor
+ */
+__kernel void depth_to_space_nchw(
+    TENSOR3D_DECLARATION(input),
+    const int batch_id,
+    TENSOR4D_DECLARATION(output))
+{
+    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(input);
+    Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0);
+
+    const int r = (CHANNEL_SIZE / (BLOCK_SHAPE * BLOCK_SHAPE));
+    const int x = get_global_id(0);
+    const int y = get_global_id(1);
+    const int z = get_global_id(2) % r;
+
+    const int out_x = x * BLOCK_SHAPE + (get_global_id(2) / r) % BLOCK_SHAPE;
+    const int out_y = y * BLOCK_SHAPE + (get_global_id(2) / r) / BLOCK_SHAPE;
+
+    *((__global DATA_TYPE *)tensor4D_offset(&out, out_x, out_y, z, batch_id)) = *((__global DATA_TYPE *)in.ptr);
+}
+/** Batch to space transformation. (NHWC)
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
+ * @note The input tensor batch size must be passed at compile time using -DCHANNEL_SIZE. e.g. -DCHANNEL_SIZE=2
+ * @note The block shape must be passed at compile time using -DBLOCK_SHAPE. e.g. -DBLOCK_SHAPE=2
+ *
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
+ * @param[in]  input_stride_x                       Stride of the source tensor 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 tensor 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 first source tensor
+ * @param[in]  batch_id                             The input tensor batch id
+ * @param[out] output_ptr                           Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination tensor 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 tensor 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 source 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 tensor
+ */
+__kernel void depth_to_space_nhwc(
+    TENSOR3D_DECLARATION(input),
+    const int batch_id,
+    TENSOR4D_DECLARATION(output))
+{
+    Tensor3D in  = CONVERT_TO_TENSOR3D_STRUCT(input);
+    Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0);
+
+    const int r = (CHANNEL_SIZE / (BLOCK_SHAPE * BLOCK_SHAPE));
+    const int x = get_global_id(1);
+    const int y = get_global_id(2);
+    const int z = get_global_id(0) % r;
+
+    const int out_x = x * BLOCK_SHAPE + (get_global_id(0) / r) % BLOCK_SHAPE;
+    const int out_y = y * BLOCK_SHAPE + (get_global_id(0) / r) / BLOCK_SHAPE;
+
+    *((__global DATA_TYPE *)tensor4D_offset(&out, z, out_x, out_y, batch_id)) = *((__global DATA_TYPE *)in.ptr);
+}
+#endif // defined(DATA_TYPE) && defined(BLOCK_SHAPE) && defined(CHANNEL_SIZE)
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLDepthToSpaceLayerKernel.cpp b/src/core/CL/kernels/CLDepthToSpaceLayerKernel.cpp
new file mode 100644
index 0000000000..f2d65e813c
--- /dev/null
+++ b/src/core/CL/kernels/CLDepthToSpaceLayerKernel.cpp
@@ -0,0 +1,128 @@
+/*
+ * Copyright (c) 2019 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/CLDepthToSpaceLayerKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+using namespace arm_compute::misc::shape_calculator;
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, int32_t block_shape)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
+    ARM_COMPUTE_RETURN_ERROR_ON(block_shape < 2);
+
+    const DataLayout data_layout = input->data_layout();
+    const int        idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_channel] % (block_shape * block_shape) != 0);
+
+    // Validate output if initialized
+    if(output->total_size() != 0)
+    {
+        const int idx_width  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+        const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+        ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_width] != (block_shape * input->tensor_shape()[idx_width]));
+        ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_height] != (block_shape * input->tensor_shape()[idx_height]));
+        ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 4);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+} // namespace
+
+CLDepthToSpaceLayerKernel::CLDepthToSpaceLayerKernel()
+    : _input(nullptr), _output(nullptr), _block_shape()
+{
+}
+
+void CLDepthToSpaceLayerKernel::configure(const ICLTensor *input, ICLTensor *output, int32_t block_shape)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    TensorShape output_shape = compute_depth_to_space_shape(input->info(), block_shape);
+    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type());
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), block_shape));
+
+    _input       = input;
+    _output      = output;
+    _block_shape = block_shape;
+
+    const int idx_width   = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
+    const int 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("-DCHANNEL_SIZE=" + support::cpp11::to_string(input->info()->dimension(idx_channel)));
+    build_opts.add_option("-DBLOCK_SHAPE=" + support::cpp11::to_string(block_shape));
+    build_opts.add_option("-DWIDTH_IN=" + support::cpp11::to_string(input->info()->dimension(idx_width)));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depth_to_space_" + lower_string(string_from_data_layout(input->info()->data_layout())), build_opts.options()));
+
+    // Configure kernel window
+    Window win = calculate_max_window(*input->info(), Steps());
+    ICLKernel::configure_internal(win);
+}
+
+Status CLDepthToSpaceLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, int32_t block_shape)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, block_shape));
+    return Status{};
+}
+
+void CLDepthToSpaceLayerKernel::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_in  = window.first_slice_window_3D();
+    Window slice_out = window.first_slice_window_4D();
+
+    slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
+    slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
+    slice_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
+    slice_out.set(3, Window::Dimension(0, 0, 0));
+
+    int batch_id = 0;
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice_in);
+        add_argument(idx, batch_id);
+        add_4D_tensor_argument(idx, _output, slice_out);
+        enqueue(queue, *this, slice_in);
+
+        ++batch_id;
+    }
+    while(window.slide_window_slice_3D(slice_in));
+}
+} // namespace arm_compute