This commit makes to support DepthToSpace op for acl neon.
- Introduce NEDepthToSpaceLayerEx and NEDepthToSpaceLayerKernelEx
- Apply NEDepthToSpaceLayerEx layer for neurun
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
--- /dev/null
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
+ * 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.
+ */
+#ifndef __ARM_COMPUTE_NEDEPTHTOSPACELAYERKERNELEX_H__
+#define __ARM_COMPUTE_NEDEPTHTOSPACELAYERKERNELEX_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Interface for the depth to space kernel */
+class NEDepthToSpaceLayerKernelEx : public INEKernel
+{
+public:
+ const char *name() const override { return "NEDepthToSpaceLayerKernelEx"; }
+ /** Default constructor */
+ NEDepthToSpaceLayerKernelEx();
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ NEDepthToSpaceLayerKernelEx(const NEDepthToSpaceLayerKernelEx &) = delete;
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ NEDepthToSpaceLayerKernelEx &operator=(const NEDepthToSpaceLayerKernelEx &) = delete;
+ /** Allow instances of this class to be moved */
+ NEDepthToSpaceLayerKernelEx(NEDepthToSpaceLayerKernelEx &&) = default;
+ /** Allow instances of this class to be moved */
+ NEDepthToSpaceLayerKernelEx &operator=(NEDepthToSpaceLayerKernelEx &&) = default;
+ /** Default destructor */
+ ~NEDepthToSpaceLayerKernelEx() = default;
+ /** Initialise the kernel's inputs and output.
+ *
+ * @param[in] input Tensor input. Supported tensor rank: 4. Data types supported:
+ * U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32.
+ * @param[out] output Tensor output. Data types supported: same as @p input
+ * @param[in] block_shape Block shape x value.
+ */
+ void configure(const ITensor *input, ITensor *output, int32_t block_shape);
+ /** Static function to check if given info will lead to a valid configuration of @ref
+ * NEDepthToSpaceLayerKernelEx.
+ *
+ * @param[in] input Tensor input info. Supported tensor rank: 4. Data types supported:
+ * U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32.
+ * @param[in] output Tensor output info. Data types supported: same as @p input
+ * @param[in] block_shape Block shape value.
+ *
+ * @return a status
+ */
+ static Status validate(const ITensorInfo *input, const ITensorInfo *output, int32_t block_shape);
+
+ // Inherited methods overridden:
+ void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+ const ITensor *_input; /**< Source tensor */
+ ITensor *_output; /**< Destination tensor */
+ int32_t _block_shape; /**< Block shape */
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NEDEPTHTOSPACELAYERKERNELEX_H__ */
return out_shape;
}
+/** Calculate the depth to space output shape of a tensor
+ *
+ * @param[in] input Input tensor info
+ * @param[in] block Block shape value
+ *
+ * @return the calculated shape
+ */
+inline TensorShape compute_depth_to_space_shape_ex(const ITensorInfo *input, int block)
+{
+ ARM_COMPUTE_ERROR_ON(block < 2);
+
+ 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);
+
+ TensorShape output_shape{input->tensor_shape()};
+ output_shape.set(idx_width, input->dimension(idx_width) * block);
+ output_shape.set(idx_height, input->dimension(idx_height) * block);
+ output_shape.set(idx_channel, input->dimension(idx_channel) / (block * block));
+
+ return output_shape;
+}
+
} // namespace shape_calculator
} // namespace misc
} // namespace arm_compute
#include <arm_compute/runtime/NEON/functions/NEArgMinMax.h>
#include <arm_compute/runtime/NEON/functions/NEBinaryLogicalOperation.h>
+#include <arm_compute/runtime/NEON/functions/NEDepthToSpaceLayerEx.h>
#include <arm_compute/runtime/NEON/functions/NEElementwiseUnaryLayerEx.h>
#include <arm_compute/runtime/NEON/functions/NEFullyConnectedReshapingLayer.h>
#include <arm_compute/runtime/NEON/functions/NEPReLU.h>
--- /dev/null
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
+ * 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.
+ */
+#ifndef __ARM_COMPUTE_NEDEPTHTOSPACELAYEREX_H__
+#define __ARM_COMPUTE_NEDEPTHTOSPACELAYEREX_H__
+
+#include "arm_compute/runtime/IFunction.h"
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/INESimpleFunctionNoBorder.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Basic function to run @ref NEDepthToSpaceLayerKernelEx. */
+class NEDepthToSpaceLayerEx : public INESimpleFunctionNoBorder
+{
+public:
+ /** Set the input and output tensors.
+ *
+ * @param[in] input Tensor input. Supported tensor rank: 4. Data types supported:
+ * U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32.
+ * @param[out] output Tensor output. Data types supported: same as @p input
+ * @param[in] block_shape Block shape value.
+ */
+ void configure(const ITensor *input, ITensor *output, int32_t block_shape);
+ /** Static function to check if given info will lead to a valid configuration of @ref
+ * NEDepthToSpaceLayerEx.
+ *
+ * @param[in] input Tensor input info. Supported tensor rank: 4. Data types supported:
+ * U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32.
+ * @param[in] output Tensor output info. Data types supported: same as @p input
+ * @param[in] block_shape Block shape x value.
+ *
+ * @return a status
+ */
+ static Status validate(const ITensorInfo *input, const ITensorInfo *output, int32_t block_shape);
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NEDEPTHTOSPACELAYEREX_H__ */
--- /dev/null
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
+ * 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/NEON/kernels/NEDepthToSpaceLayerKernelEx.h"
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/wrapper/wrapper.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculatorEx.h"
+#include <arm_neon.h>
+#include <cstdint>
+
+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
+
+NEDepthToSpaceLayerKernelEx::NEDepthToSpaceLayerKernelEx()
+ : _input(nullptr), _output(nullptr), _block_shape()
+{
+}
+
+void NEDepthToSpaceLayerKernelEx::configure(const ITensor *input, ITensor *output,
+ int32_t block_shape)
+{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+ TensorShape output_shape = compute_depth_to_space_shape_ex(input->info(), block_shape);
+ // Output auto inizialitation if not yet initialized
+ auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
+
+ // Perform validation step
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), block_shape));
+
+ _input = input;
+ _output = output;
+ _block_shape = block_shape;
+
+ // Configure kernel window
+ Window win = calculate_max_window(*input->info(), Steps());
+ ICPPKernel::configure(win);
+}
+
+Status NEDepthToSpaceLayerKernelEx::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 NEDepthToSpaceLayerKernelEx::run(const Window &window, const ThreadInfo &info)
+{
+ ARM_COMPUTE_UNUSED(info);
+ ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+ ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICPPKernel::window(), window);
+
+ const int idx_channel =
+ get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::CHANNEL);
+ const int depth_size = _input->info()->dimension(idx_channel);
+ const int r = (depth_size / (_block_shape * _block_shape));
+ const int element_size = _input->info()->element_size();
+
+ Window slice_out = window.first_slice_window_3D();
+
+ // The slice_out slice does not move
+ 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));
+
+ // Main loop for NCHW and NHWC
+ if (_input->info()->data_layout() == DataLayout::NCHW)
+ {
+ Window slice_in = window.first_slice_window_2D();
+ do
+ {
+ Iterator in(_input, slice_in);
+ execute_window_loop(slice_in,
+ [&](const Coordinates &id) {
+ const int x = id.x();
+ const int y = id.y();
+
+ const int z = id.z() % r;
+ const int out_x = x * _block_shape + (id.z() / r) % _block_shape;
+ const int out_y = y * _block_shape + (id.z() / r) / _block_shape;
+ Coordinates output_coords{out_x, out_y, z, id[3]};
+ memcpy(_output->ptr_to_element(output_coords), in.ptr(), element_size);
+ },
+ in);
+ } while (window.slide_window_slice_2D(slice_in));
+ }
+ else
+ {
+ Window slice_in = window.first_slice_window_3D();
+ do
+ {
+ Iterator in(_input, slice_in);
+ execute_window_loop(slice_in,
+ [&](const Coordinates &id) {
+ const int x = id.y();
+ const int y = id.z();
+
+ const int z = id.x() % r;
+ const int out_x = x * _block_shape + (id.x() / r) % _block_shape;
+ const int out_y = y * _block_shape + (id.x() / r) / _block_shape;
+ Coordinates output_coords{z, out_x, out_y, id[3]};
+ memcpy(_output->ptr_to_element(output_coords), in.ptr(), element_size);
+ },
+ in);
+ } while (window.slide_window_slice_3D(slice_in));
+ }
+}
+} // namespace arm_compute
--- /dev/null
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
+ * 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/NEON/functions/NEDepthToSpaceLayerEx.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/NEON/kernels/NEDepthToSpaceLayerKernelEx.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+
+namespace arm_compute
+{
+void NEDepthToSpaceLayerEx::configure(const ITensor *input, ITensor *output, int32_t block_shape)
+{
+ auto k = arm_compute::support::cpp14::make_unique<NEDepthToSpaceLayerKernelEx>();
+ k->configure(input, output, block_shape);
+ _kernel = std::move(k);
+}
+
+Status NEDepthToSpaceLayerEx::validate(const ITensorInfo *input, const ITensorInfo *output,
+ int32_t block_shape)
+{
+ return NEDepthToSpaceLayerKernelEx::validate(input, output, block_shape);
+}
+} // namespace arm_compute
_execution_builder->append(asAclFunction(std::move(fn)));
}
+void KernelGenerator::visit(const model::operation::DepthToSpaceNode &node)
+{
+ const auto output_index{node.getOutputs().at(0)};
+ const auto input_index{node.getInputs().at(model::operation::DepthToSpaceNode::Input::INPUT)};
+ const auto block_size_index{node.param().block_size_index};
+
+ auto block_size = _ctx.at(block_size_index).asScalar<int32_t>();
+ assert(block_size > 0);
+
+ auto output_alloc = _tensor_builder->at(output_index).get();
+ auto input_alloc = _tensor_builder->at(input_index).get();
+
+ auto fn = nnfw::cpp14::make_unique<::arm_compute::NEDepthToSpaceLayerEx>();
+
+ fn->configure(input_alloc->handle(), output_alloc->handle(), block_size);
+
+ auto acl_fn = asAclFunction(std::move(fn));
+
+ _execution_builder->append(std::move(acl_fn));
+}
+
void KernelGenerator::visit(const model::operation::DepthwiseConv2DNode &node)
{
using model::operation::DepthwiseConv2DNode;
void visit(const model::operation::AbsNode &) override;
void visit(const model::operation::ArgMaxNode &) override;
void visit(const model::operation::Conv2DNode &) override;
+ void visit(const model::operation::DepthToSpaceNode &) override;
void visit(const model::operation::DepthwiseConv2DNode &) override;
void visit(const model::operation::DequantizeNode &) override;
void visit(const model::operation::MaxPool2DNode &) override;
void ShapeFixer::visit(const model::operation::Conv2DNode &) { /* DO NOTHING */}
+void ShapeFixer::visit(const model::operation::DepthToSpaceNode &) { /* DO NOTHING */}
+
void ShapeFixer::visit(const model::operation::DepthwiseConv2DNode &) { /* DO NOTHING */}
void ShapeFixer::visit(const model::operation::DequantizeNode &) { /* DO NOTHING */}
void visit(const model::operation::AbsNode &) override;
void visit(const model::operation::ArgMaxNode &) override;
void visit(const model::operation::Conv2DNode &) override;
+ void visit(const model::operation::DepthToSpaceNode &) override;
void visit(const model::operation::DepthwiseConv2DNode &) override;
void visit(const model::operation::DequantizeNode &) override;
void visit(const model::operation::MaxPool2DNode &) override;
#
# Not support operations
TrivialTest.BroadcastMulTwo
-GeneratedTests.depth_to_space*
GeneratedTests.dequantize
GeneratedTests.embedding_lookup*
GeneratedTests.hashtable_lookup*
projects / platform / core / ml / nnfw.git / commitdiff