diff options
author | Michalis Spyrou <michalis.spyrou@arm.com> | 2019-05-29 17:03:38 +0100 |
---|---|---|
committer | Michalis Spyrou <michalis.spyrou@arm.com> | 2019-06-06 15:30:10 +0000 |
commit | d69b3b2db4b359d4e68cf786a294fa1d7d4db2d0 (patch) | |
tree | f527b3e7aeb6b1b565fb8ef921250cc50ca0457a /src | |
parent | fad18382a77655b8139c5c82b86ea34bc642a27a (diff) | |
download | ComputeLibrary-d69b3b2db4b359d4e68cf786a294fa1d7d4db2d0.tar.gz |
COMPMID-2238: Implement SPACE_TO_DEPTH for CL
Change-Id: I2f971d6902a64ee40820b6959e9b0b4224da6947
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1283
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src')
-rw-r--r-- | src/core/CL/CLKernelLibrary.cpp | 6 | ||||
-rw-r--r-- | src/core/CL/cl_kernels/space_to_depth.cl | 111 | ||||
-rw-r--r-- | src/core/CL/kernels/CLSpaceToDepthLayerKernel.cpp | 129 | ||||
-rw-r--r-- | src/runtime/CL/functions/CLSpaceToDepthLayer.cpp | 54 |
4 files changed, 300 insertions, 0 deletions
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp index 9806796203..4c9808d7a4 100644 --- a/src/core/CL/CLKernelLibrary.cpp +++ b/src/core/CL/CLKernelLibrary.cpp @@ -470,6 +470,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map = { "space_to_batch_static_nchw", "space_to_batch.cl" }, { "space_to_batch_nhwc", "space_to_batch.cl" }, { "space_to_batch_static_nhwc", "space_to_batch.cl" }, + { "space_to_depth_nchw", "space_to_depth.cl" }, + { "space_to_depth_nhwc", "space_to_depth.cl" }, { "softmax_layer_max_shift_exp_sum_parallel", "softmax_layer.cl" }, { "stack_layer", "stack_layer.cl" }, { "strided_slice", "slice_ops.cl" }, @@ -959,6 +961,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map = #include "./cl_kernels/space_to_batch.clembed" }, { + "space_to_depth.cl", +#include "./cl_kernels/space_to_depth.clembed" + }, + { "stack_layer.cl", #include "./cl_kernels/stack_layer.clembed" }, diff --git a/src/core/CL/cl_kernels/space_to_depth.cl b/src/core/CL/cl_kernels/space_to_depth.cl new file mode 100644 index 0000000000..898287efaa --- /dev/null +++ b/src/core/CL/cl_kernels/space_to_depth.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) +/** Space to depth 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 space_to_depth_nchw( + TENSOR4D_DECLARATION(input), + const int batch_id, + TENSOR3D_DECLARATION(output)) +{ + Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0); + Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output); + + 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 in_x = x * BLOCK_SHAPE + (get_global_id(2) / r) % BLOCK_SHAPE; + const int in_y = y * BLOCK_SHAPE + (get_global_id(2) / r) / BLOCK_SHAPE; + + *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, in_x, in_y, z, batch_id)); +} +/** Space to depth 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 space_to_depth_nhwc( + TENSOR4D_DECLARATION(input), + const int batch_id, + TENSOR3D_DECLARATION(output)) +{ + Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0); + Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output); + + 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 in_x = x * BLOCK_SHAPE + (get_global_id(0) / r) % BLOCK_SHAPE; + const int in_y = y * BLOCK_SHAPE + (get_global_id(0) / r) / BLOCK_SHAPE; + + *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, z, in_x, in_y, batch_id)); +} +#endif // defined(DATA_TYPE) && defined(BLOCK_SHAPE) && defined(CHANNEL_SIZE)
\ No newline at end of file diff --git a/src/core/CL/kernels/CLSpaceToDepthLayerKernel.cpp b/src/core/CL/kernels/CLSpaceToDepthLayerKernel.cpp new file mode 100644 index 0000000000..57d8305e47 --- /dev/null +++ b/src/core/CL/kernels/CLSpaceToDepthLayerKernel.cpp @@ -0,0 +1,129 @@ +/* + * 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/CLSpaceToDepthLayerKernel.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 < 1); + + // Validate output if initialized + if(output->total_size() != 0) + { + const DataLayout data_layout = input->data_layout(); + 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); + const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); + const int idx_batch = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); + ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_width] % block_shape != 0); + ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_height] % block_shape != 0); + ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_batch] != output->tensor_shape()[idx_batch]); + ARM_COMPUTE_RETURN_ERROR_ON(output->tensor_shape()[idx_channel] % (block_shape * block_shape) != 0); + ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape().total_size() != output->tensor_shape().total_size()); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + } + + return Status{}; +} +} // namespace + +CLSpaceToDepthLayerKernel::CLSpaceToDepthLayerKernel() + : _input(nullptr), _output(nullptr), _block_shape() +{ +} + +void CLSpaceToDepthLayerKernel::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(output->info()->data_type())); + build_opts.add_option("-DCHANNEL_SIZE=" + support::cpp11::to_string(output->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(output->info()->dimension(idx_width))); + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("space_to_depth_" + lower_string(string_from_data_layout(input->info()->data_layout())), build_opts.options())); + + // Configure kernel window + Window win = calculate_max_window(*output->info(), Steps()); + ICLKernel::configure_internal(win); +} + +Status CLSpaceToDepthLayerKernel::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 CLSpaceToDepthLayerKernel::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_out = window.first_slice_window_3D(); + Window slice_in = window.first_slice_window_4D(); + + slice_in.set(Window::DimX, Window::Dimension(0, 0, 0)); + slice_in.set(Window::DimY, Window::Dimension(0, 0, 0)); + slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0)); + slice_in.set(3, Window::Dimension(0, 0, 0)); + + int batch_id = 0; + do + { + unsigned int idx = 0; + add_4D_tensor_argument(idx, _input, slice_in); + add_argument(idx, batch_id); + add_3D_tensor_argument(idx, _output, slice_out); + enqueue(queue, *this, slice_out); + + ++batch_id; + } + while(window.slide_window_slice_3D(slice_out)); +} +} // namespace arm_compute diff --git a/src/runtime/CL/functions/CLSpaceToDepthLayer.cpp b/src/runtime/CL/functions/CLSpaceToDepthLayer.cpp new file mode 100644 index 0000000000..f02a13f66d --- /dev/null +++ b/src/runtime/CL/functions/CLSpaceToDepthLayer.cpp @@ -0,0 +1,54 @@ +/* + * 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/runtime/CL/functions/CLSpaceToDepthLayer.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/runtime/CL/CLScheduler.h" + +namespace arm_compute +{ +CLSpaceToDepthLayer::CLSpaceToDepthLayer() + : _space_to_depth_kernel() +{ +} + +void CLSpaceToDepthLayer::configure(const ICLTensor *input, ICLTensor *output, int32_t block_shape) +{ + _space_to_depth_kernel.configure(input, output, block_shape); +} + +Status CLSpaceToDepthLayer::validate(const ITensorInfo *input, const ITensorInfo *output, int32_t block_shape) +{ + return CLSpaceToDepthLayerKernel::validate(input, output, block_shape); +} + +void CLSpaceToDepthLayer::run() +{ + CLScheduler::get().enqueue(_space_to_depth_kernel, true); +} +} // namespace arm_compute |