From 7c9541ccd4c98d7e9a456ee67c3ceecce8531ffb Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Thu, 20 Sep 2018 17:40:04 +0100 Subject: COMPMID-1596 Create UpsampleLayer for NEON Change-Id: I82d95c4f1c5fed13b213a2591cc2b4e0d0e02a54 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/149676 Reviewed-by: Georgios Pinitas Reviewed-by: Pablo Tello Tested-by: bsgcomp --- arm_compute/core/NEON/NEKernels.h | 1 + .../core/NEON/kernels/NEUpsampleLayerKernel.h | 121 +++++++ arm_compute/runtime/NEON/NEFunctions.h | 1 + .../runtime/NEON/functions/NEUpsampleLayer.h | 73 ++++ src/core/NEON/kernels/NEUpsampleLayerKernel.cpp | 376 +++++++++++++++++++++ src/graph/backends/NEON/NEFunctionFactory.cpp | 2 + src/graph/backends/NEON/NENodeValidator.cpp | 2 +- src/runtime/NEON/functions/NEUpsampleLayer.cpp | 52 +++ tests/validation/NEON/Upsample.cpp | 156 +++++++++ tests/validation/fixtures/UpsampleLayerFixture.h | 14 +- tests/validation/reference/UpsampleLayer.cpp | 12 +- tests/validation/reference/UpsampleLayer.h | 2 +- 12 files changed, 798 insertions(+), 14 deletions(-) create mode 100644 arm_compute/core/NEON/kernels/NEUpsampleLayerKernel.h create mode 100644 arm_compute/runtime/NEON/functions/NEUpsampleLayer.h create mode 100644 src/core/NEON/kernels/NEUpsampleLayerKernel.cpp create mode 100644 src/runtime/NEON/functions/NEUpsampleLayer.cpp create mode 100644 tests/validation/NEON/Upsample.cpp diff --git a/arm_compute/core/NEON/NEKernels.h b/arm_compute/core/NEON/NEKernels.h index 2b76c0bd29..cb2e851b66 100644 --- a/arm_compute/core/NEON/NEKernels.h +++ b/arm_compute/core/NEON/NEKernels.h @@ -113,6 +113,7 @@ #include "arm_compute/core/NEON/kernels/NETableLookupKernel.h" #include "arm_compute/core/NEON/kernels/NEThresholdKernel.h" #include "arm_compute/core/NEON/kernels/NETransposeKernel.h" +#include "arm_compute/core/NEON/kernels/NEUpsampleLayerKernel.h" #include "arm_compute/core/NEON/kernels/NEWarpKernel.h" #include "arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h" #include "arm_compute/core/NEON/kernels/NEWidthConcatenateLayerKernel.h" diff --git a/arm_compute/core/NEON/kernels/NEUpsampleLayerKernel.h b/arm_compute/core/NEON/kernels/NEUpsampleLayerKernel.h new file mode 100644 index 0000000000..cf21e6cf73 --- /dev/null +++ b/arm_compute/core/NEON/kernels/NEUpsampleLayerKernel.h @@ -0,0 +1,121 @@ +/* + * 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. + */ +#ifndef __ARM_COMPUTE_NEUPSAMPLELAYERKERNEL_H__ +#define __ARM_COMPUTE_NEUPSAMPLELAYERKERNEL_H__ + +#include "arm_compute/core/NEON/INEKernel.h" + +namespace arm_compute +{ +class ITensor; + +/** Interface for the Upsample layer kernel.*/ +class NEUpsampleLayerKernel : public INEKernel +{ +public: + const char *name() const override + { + return "NEUpsampleLayerKernel"; + } + /** Default constructor */ + NEUpsampleLayerKernel(); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEUpsampleLayerKernel(const NEUpsampleLayerKernel &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEUpsampleLayerKernel &operator=(const NEUpsampleLayerKernel &) = delete; + /** Default Move Constructor. */ + NEUpsampleLayerKernel(NEUpsampleLayerKernel &&) = default; + /** Default move assignment operator */ + NEUpsampleLayerKernel &operator=(NEUpsampleLayerKernel &&) = default; + /** Default destructor */ + ~NEUpsampleLayerKernel() = default; + /** Set the input output tensors. + * + * @param[in] input Source tensor. Data types supported: QASYMM8/F16/F32. + * @param[out] output Destination tensor. Data types supported: same as @p input. + * @param[in] info Contains stride information described in @ref Size2D. + * @param[in] policy Defines the policy to fill the intermediate pixels. + * + */ + void configure(const ITensor *input, ITensor *output, const Size2D &info, const InterpolationPolicy policy); + /** Static function to check if given info will lead to a valid configuration of @ref NEUpsampleLayerKernel + * + * @param[in] input Source tensor info. Data types supported: QASYMM8/F16/F32. + * @param[in] output Destination tensor info. Data types supported: same as @p input. + * @param[in] info Contains stride information described in @ref Size2D. + * @param[in] policy Defines the policy to fill the intermediate pixels. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &info, const InterpolationPolicy policy); + + // Inherited methods overridden: + void run(const Window &window, const ThreadInfo &info) override; + +private: + /** Function to run upsample layer for FP32 (NCHW) + * + * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). + */ + void upsample_f32_nchw(const Window &window); + /** Function to run upsample layer for FP32 (NHWC) + * + * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). + */ + void upsample_f32_nhwc(const Window &window); + /** Function to run upsample layer for FP16 (NCHW) + * + * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). + */ + void upsample_f16_nchw(const Window &window); + /** Function to run upsample layer for FP16 (NHWC) + * + * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). + */ + void upsample_f16_nhwc(const Window &window); + /** Function to run upsample layer for QASYMM8 (NCHW) + * + * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). + */ + void upsample_qasymm8_nchw(const Window &window); + /** Function to run upsample layer for QASYMM8 (NHWC) + * + * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). + */ + void upsample_qasymm8_nhwc(const Window &window); + /** Common signature for all the upsample layer functions + * + * @param[in] window Region on which to execute the kernel. + */ + using UpsampleFunctionPtr = void (NEUpsampleLayerKernel::*)(const Window &window); + +private: + UpsampleFunctionPtr _func; + const ITensor *_input; + ITensor *_output; + Size2D _info; + unsigned int _num_elems_processed_per_iteration_x; +}; +} // namespace arm_compute +#endif /*__ARM_COMPUTE_NEUPSAMPLELAYERKERNEL_H__ */ diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h index d9b7269efc..2bf8bcd515 100644 --- a/arm_compute/runtime/NEON/NEFunctions.h +++ b/arm_compute/runtime/NEON/NEFunctions.h @@ -115,6 +115,7 @@ #include "arm_compute/runtime/NEON/functions/NETableLookup.h" #include "arm_compute/runtime/NEON/functions/NEThreshold.h" #include "arm_compute/runtime/NEON/functions/NETranspose.h" +#include "arm_compute/runtime/NEON/functions/NEUpsampleLayer.h" #include "arm_compute/runtime/NEON/functions/NEWarpAffine.h" #include "arm_compute/runtime/NEON/functions/NEWarpPerspective.h" #include "arm_compute/runtime/NEON/functions/NEWidthConcatenateLayer.h" diff --git a/arm_compute/runtime/NEON/functions/NEUpsampleLayer.h b/arm_compute/runtime/NEON/functions/NEUpsampleLayer.h new file mode 100644 index 0000000000..b88e71c0d7 --- /dev/null +++ b/arm_compute/runtime/NEON/functions/NEUpsampleLayer.h @@ -0,0 +1,73 @@ +/* + * 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. + */ +#ifndef __ARM_COMPUTE_NEUPSAMPLELAYER_H__ +#define __ARM_COMPUTE_NEUPSAMPLELAYER_H__ + +#include "arm_compute/core/NEON/kernels/NEUpsampleLayerKernel.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/IFunction.h" +#include "arm_compute/runtime/NEON/NEScheduler.h" +#include "arm_compute/runtime/Tensor.h" + +namespace arm_compute +{ +class ITensor; + +/** Function to run upsample layer */ +class NEUpsampleLayer : public IFunction +{ +public: + /** Constructor */ + NEUpsampleLayer(); + /** Set the input output tensors. + * + * @param[in] input Source tensor. Data types supported: QASYMM8/F16/F32. + * @param[out] output Destination tensor. Data types supported: same as @p input. + * @param[in] info Contains stride information described in @ref Size2D. + * @param[in] policy Defines the policy to fill the intermediate pixels. + * + */ + void configure(const ITensor *input, ITensor *output, const Size2D &info, + const InterpolationPolicy &policy); + /** Static function to check if given info will lead to a valid configuration of @ref NEUpsampleLayer + * + * @param[in] input Source tensor info. Data types supported: QASYMM8/F16/F32. + * @param[out] output Destination tensor info. Data types supported: same as @p input. + * @param[in] info Contains stride information described in @ref Size2D. + * @param[in] policy Defines the policy to fill the intermediate pixels. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &info, + const InterpolationPolicy &policy); + + // Inherited methods overridden: + void run() override; + +private: + NEUpsampleLayerKernel _kernel; + DataLayout _data_layout; +}; +} // arm_compute +#endif /* __ARM_COMPUTE_NEUPSAMPLELAYER_H__ */ diff --git a/src/core/NEON/kernels/NEUpsampleLayerKernel.cpp b/src/core/NEON/kernels/NEUpsampleLayerKernel.cpp new file mode 100644 index 0000000000..5dca58edd8 --- /dev/null +++ b/src/core/NEON/kernels/NEUpsampleLayerKernel.cpp @@ -0,0 +1,376 @@ +/* + * 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/NEON/kernels/NEUpsampleLayerKernel.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.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 + +namespace arm_compute +{ +namespace +{ +std::pair validate_and_configure_window_nchw(ITensorInfo *input, ITensorInfo *output, int num_elems_processed_per_iteration_x, const Size2D &info) +{ + const int num_elems_processed_per_iteration_x_out = num_elems_processed_per_iteration_x * info.x(); + Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x_out)); + AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, 1, 0.5f, 0.5f); + AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration_x_out); + bool window_changed = update_window_and_padding(win, input_access, output_access); + output_access.set_valid_region(win, output->valid_region()); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} + +std::pair validate_and_configure_window_nhwc(ITensorInfo *input, ITensorInfo *output, int num_elems_processed_per_iteration_x, const Size2D &info) +{ + ARM_COMPUTE_UNUSED(info); + Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x)); + AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration_x); + AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration_x); + bool window_changed = update_window_and_padding(win, input_access, output_access); + output_access.set_valid_region(win, output->valid_region()); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} + +std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, int num_elems_processed_per_iteration_x, const Size2D &info) +{ + std::pair win_config; + switch(input->data_layout()) + { + case DataLayout::NCHW: + win_config = validate_and_configure_window_nchw(input, output, num_elems_processed_per_iteration_x, info); + break; + case DataLayout::NHWC: + win_config = validate_and_configure_window_nhwc(input, output, num_elems_processed_per_iteration_x, info); + break; + default: + win_config = std::make_pair(ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported data layout!"), Window{}); + } + + return win_config; +} +} // namespace +NEUpsampleLayerKernel::NEUpsampleLayerKernel() + : _func(nullptr), _input(nullptr), _output(nullptr), _info(), _num_elems_processed_per_iteration_x() +{ +} + +Status NEUpsampleLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &info, const InterpolationPolicy policy) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_UNUSED(policy); + + 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); + + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.x() != 2 || info.y() != 2, "Only stride 2 is supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(policy != InterpolationPolicy::NEAREST_NEIGHBOR, "Only nearest neighbor policy supported"); + + // Check output if configured + 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(idx_width) != info.x() * input->dimension(idx_width)); + ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(idx_height) != info.y() * input->dimension(idx_height)); + } + + const int num_elems_processed_per_iteration_x = 16 / input->element_size(); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), + output->clone().get(), num_elems_processed_per_iteration_x, info) + .first); + return Status{}; +} + +void NEUpsampleLayerKernel::upsample_f32_nchw(const arm_compute::Window &window) +{ + Window window_in(window); + window_in.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _num_elems_processed_per_iteration_x)); + + Window window_out(window); + window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _info.y())); + + Iterator input(_input, window_in); + Iterator output(_output, window_out); + const int offset_y_out = _output->info()->strides_in_bytes().y() / sizeof(float); + + execute_window_loop(window_out, [&](const Coordinates & id) + { + const float32x4_t data = vld1q_f32(reinterpret_cast(input.ptr())); + const float32x4_t data_out1 = { vgetq_lane_f32(data, 0), vgetq_lane_f32(data, 0), vgetq_lane_f32(data, 1), vgetq_lane_f32(data, 1) }; + const float32x4_t data_out2 = { vgetq_lane_f32(data, 2), vgetq_lane_f32(data, 2), vgetq_lane_f32(data, 3), vgetq_lane_f32(data, 3) }; + auto out = reinterpret_cast(output.ptr()); + + vst1q_f32(out, data_out1); + vst1q_f32(out + 4, data_out2); + vst1q_f32(out + offset_y_out, data_out1); + vst1q_f32(out + offset_y_out + 4, data_out2); + }, + input, output); +} + +void NEUpsampleLayerKernel::upsample_f32_nhwc(const arm_compute::Window &window) +{ + Window window_out(window); + window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _info.x())); + window_out.set(Window::DimZ, Window::Dimension(0, _output->info()->dimension(2), _info.y())); + + Iterator input(_input, window); + Iterator output(_output, window_out); + + const int offset_y_out = _output->info()->strides_in_bytes().y() / sizeof(float); + const int offset_z_out = _output->info()->strides_in_bytes().z() / sizeof(float); + + execute_window_loop(window_out, [&](const Coordinates & id) + { + const float32x4_t data = vld1q_f32(reinterpret_cast(input.ptr())); + auto out = reinterpret_cast(output.ptr()); + + vst1q_f32(out, data); + vst1q_f32(out + offset_y_out, data); + vst1q_f32(out + offset_z_out, data); + vst1q_f32(out + offset_y_out + offset_z_out, data); + }, + input, output); +} + +void NEUpsampleLayerKernel::upsample_qasymm8_nchw(const arm_compute::Window &window) +{ + Window window_in(window); + window_in.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _num_elems_processed_per_iteration_x)); + + Window window_out(window); + window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _info.y())); + + Iterator input(_input, window_in); + Iterator output(_output, window_out); + const int offset_y_out = _output->info()->strides_in_bytes().y() / sizeof(uint8_t); + + execute_window_loop(window_out, [&](const Coordinates & id) + { + const uint8x16_t data = vld1q_u8(reinterpret_cast(input.ptr())); + const uint8x16_t data_out1 = { vgetq_lane_u8(data, 0), vgetq_lane_u8(data, 0), vgetq_lane_u8(data, 1), vgetq_lane_u8(data, 1), + vgetq_lane_u8(data, 2), vgetq_lane_u8(data, 2), vgetq_lane_u8(data, 3), vgetq_lane_u8(data, 3), + vgetq_lane_u8(data, 4), vgetq_lane_u8(data, 4), vgetq_lane_u8(data, 5), vgetq_lane_u8(data, 5), + vgetq_lane_u8(data, 6), vgetq_lane_u8(data, 6), vgetq_lane_u8(data, 7), vgetq_lane_u8(data, 7) + }; + const uint8x16_t data_out2 = + { + vgetq_lane_u8(data, 8), vgetq_lane_u8(data, 8), vgetq_lane_u8(data, 9), vgetq_lane_u8(data, 9), + vgetq_lane_u8(data, 10), vgetq_lane_u8(data, 10), vgetq_lane_u8(data, 11), vgetq_lane_u8(data, 11), + vgetq_lane_u8(data, 12), vgetq_lane_u8(data, 12), vgetq_lane_u8(data, 13), vgetq_lane_u8(data, 13), + vgetq_lane_u8(data, 14), vgetq_lane_u8(data, 14), vgetq_lane_u8(data, 15), vgetq_lane_u8(data, 15) + }; + auto out = reinterpret_cast(output.ptr()); + + vst1q_u8(out, data_out1); + vst1q_u8(out + 16, data_out2); + vst1q_u8(out + offset_y_out, data_out1); + vst1q_u8(out + offset_y_out + 16, data_out2); + }, + input, output); +} + +void NEUpsampleLayerKernel::upsample_qasymm8_nhwc(const arm_compute::Window &window) +{ + Window window_out(window); + window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _info.x())); + window_out.set(Window::DimZ, Window::Dimension(0, _output->info()->dimension(2), _info.y())); + + Iterator input(_input, window); + Iterator output(_output, window_out); + + const int offset_y_out = _output->info()->strides_in_bytes().y() / sizeof(uint8_t); + const int offset_z_out = _output->info()->strides_in_bytes().z() / sizeof(uint8_t); + execute_window_loop(window_out, [&](const Coordinates & id) + { + const uint8x16_t data = vld1q_u8(reinterpret_cast(input.ptr())); + auto out = reinterpret_cast(output.ptr()); + + vst1q_u8(out, data); + vst1q_u8(out + offset_y_out, data); + vst1q_u8(out + offset_z_out, data); + vst1q_u8(out + offset_y_out + offset_z_out, data); + }, + input, output); +} + +void NEUpsampleLayerKernel::upsample_f16_nchw(const arm_compute::Window &window) +{ + ARM_COMPUTE_UNUSED(window); +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + Window window_in(window); + window_in.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _num_elems_processed_per_iteration_x)); + + Window window_out(window); + window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _info.y())); + + Iterator input(_input, window_in); + Iterator output(_output, window_out); + const int offset_y_out = _output->info()->strides_in_bytes().y() / sizeof(float16_t); + + execute_window_loop(window_out, [&](const Coordinates & id) + { + const float16x8_t data = vld1q_f16(reinterpret_cast(input.ptr())); + const float16x8_t data_out1 = { vgetq_lane_f16(data, 0), vgetq_lane_f16(data, 0), vgetq_lane_f16(data, 1), vgetq_lane_f16(data, 1), + vgetq_lane_f16(data, 2), vgetq_lane_f16(data, 2), vgetq_lane_f16(data, 3), vgetq_lane_f16(data, 3) + }; + const float16x8_t data_out2 = { vgetq_lane_f16(data, 4), vgetq_lane_f16(data, 4), vgetq_lane_f16(data, 5), vgetq_lane_f16(data, 5), + vgetq_lane_f16(data, 6), vgetq_lane_f16(data, 6), vgetq_lane_f16(data, 7), vgetq_lane_f16(data, 7) + }; + auto out = reinterpret_cast(output.ptr()); + + vst1q_f16(out, data_out1); + vst1q_f16(out + 8, data_out2); + vst1q_f16(out + offset_y_out, data_out1); + vst1q_f16(out + offset_y_out + 8, data_out2); + }, + input, output); +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +} + +void NEUpsampleLayerKernel::upsample_f16_nhwc(const arm_compute::Window &window) +{ + ARM_COMPUTE_UNUSED(window); +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + Window window_out(window); + window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _info.x())); + window_out.set(Window::DimZ, Window::Dimension(0, _output->info()->dimension(2), _info.y())); + + Iterator input(_input, window); + Iterator output(_output, window_out); + const int offset_y_out = _output->info()->strides_in_bytes().y() / sizeof(float16_t); + const int offset_z_out = _output->info()->strides_in_bytes().z() / sizeof(float16_t); + + execute_window_loop(window_out, [&](const Coordinates & id) + { + const float16x8_t data = vld1q_f16(reinterpret_cast(input.ptr())); + auto out = reinterpret_cast(output.ptr()); + + vst1q_f16(out, data); + vst1q_f16(out + offset_y_out, data); + vst1q_f16(out + offset_z_out, data); + vst1q_f16(out + offset_y_out + offset_z_out, data); + }, + input, output); +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +} + +void NEUpsampleLayerKernel::configure(const ITensor *input, ITensor *output, const Size2D &info, const InterpolationPolicy policy) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_UNUSED(policy); + + _input = input; + _output = output; + _info = info; + + const DataLayout data_layout = input->info()->data_layout(); + + TensorShape output_shape = misc::shape_calculator::compute_upsample_shape(*input->info(), info); + auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type()); + output->info()->set_data_layout(data_layout); + + // Perform validation step + ARM_COMPUTE_ERROR_THROW_ON(NEUpsampleLayerKernel::validate(input->info(), output->info(), info, policy)); + + _num_elems_processed_per_iteration_x = 16 / output->info()->element_size(); + + switch(data_layout) + { + case DataLayout::NCHW: + { + switch(input->info()->data_type()) + { + case DataType::QASYMM8: + _func = &NEUpsampleLayerKernel::upsample_qasymm8_nchw; + break; + case DataType::F32: + _func = &NEUpsampleLayerKernel::upsample_f32_nchw; + break; +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + case DataType::F16: + _func = &NEUpsampleLayerKernel::upsample_f16_nchw; + break; +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + default: + ARM_COMPUTE_ERROR("Not implemented"); + } + break; + } + case DataLayout::NHWC: + { + switch(input->info()->data_type()) + { + case DataType::QASYMM8: + _func = &NEUpsampleLayerKernel::upsample_qasymm8_nhwc; + break; + case DataType::F32: + _func = &NEUpsampleLayerKernel::upsample_f32_nhwc; + break; +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + case DataType::F16: + _func = &NEUpsampleLayerKernel::upsample_f16_nhwc; + break; +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + default: + ARM_COMPUTE_ERROR("Not implemented"); + } + break; + } + default: + ARM_COMPUTE_ERROR("Not implemented"); + } + + // Configure window + std::pair win_config = validate_and_configure_window(input->info(), + output->info(), + _num_elems_processed_per_iteration_x, + info); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + INEKernel::configure(win_config.second); +} + +void NEUpsampleLayerKernel::run(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); + ARM_COMPUTE_ERROR_ON(_func == nullptr); + + (this->*_func)(window); +} +} // namespace arm_compute diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp index ec70a5ae93..508e4e8ec4 100644 --- a/src/graph/backends/NEON/NEFunctionFactory.cpp +++ b/src/graph/backends/NEON/NEFunctionFactory.cpp @@ -215,6 +215,8 @@ std::unique_ptr NEFunctionFactory::create(INode *node, GraphContext & return detail::create_resize_layer(*polymorphic_downcast(node)); case NodeType::SoftmaxLayer: return detail::create_softmax_layer(*polymorphic_downcast(node), ctx); + case NodeType::UpsampleLayer: + return detail::create_upsample_layer(*polymorphic_downcast(node), ctx); default: return nullptr; } diff --git a/src/graph/backends/NEON/NENodeValidator.cpp b/src/graph/backends/NEON/NENodeValidator.cpp index 4f35298c7d..fd3e6f5391 100644 --- a/src/graph/backends/NEON/NENodeValidator.cpp +++ b/src/graph/backends/NEON/NENodeValidator.cpp @@ -68,7 +68,7 @@ Status NENodeValidator::validate(INode *node) case NodeType::SliceLayer: return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : SliceLayer"); case NodeType::UpsampleLayer: - return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : UpsampleLayer"); + return detail::validate_upsample_layer(*polymorphic_downcast(node)); case NodeType::YOLOLayer: return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : YOLOLayer"); default: diff --git a/src/runtime/NEON/functions/NEUpsampleLayer.cpp b/src/runtime/NEON/functions/NEUpsampleLayer.cpp new file mode 100644 index 0000000000..9be96af66a --- /dev/null +++ b/src/runtime/NEON/functions/NEUpsampleLayer.cpp @@ -0,0 +1,52 @@ +/* + * 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/runtime/NEON/functions/NEUpsampleLayer.h" + +#include "arm_compute/core/NEON/kernels/NEUpsampleLayerKernel.h" + +namespace arm_compute +{ +NEUpsampleLayer::NEUpsampleLayer() + : _kernel(), _data_layout() +{ +} + +Status NEUpsampleLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &info, + const InterpolationPolicy &policy) +{ + return NEUpsampleLayerKernel::validate(input, output, info, policy); +} + +void NEUpsampleLayer::configure(const ITensor *input, ITensor *output, const Size2D &info, const InterpolationPolicy &policy) +{ + _data_layout = input->info()->data_layout(); + _kernel.configure(input, output, info, policy); +} + +void NEUpsampleLayer::run() +{ + const auto win = (_data_layout == DataLayout::NCHW) ? Window::DimZ : Window::DimX; + NEScheduler::get().schedule(&_kernel, win); +} +} // namespace arm_compute diff --git a/tests/validation/NEON/Upsample.cpp b/tests/validation/NEON/Upsample.cpp new file mode 100644 index 0000000000..39b69ee1e3 --- /dev/null +++ b/tests/validation/NEON/Upsample.cpp @@ -0,0 +1,156 @@ +/* + * 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/Types.h" +#include "arm_compute/runtime/NEON/functions/NEUpsampleLayer.h" +#include "tests/NEON/Accessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/UpsampleLayerFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +TEST_SUITE(NEON) +TEST_SUITE(UpsampleLayer) + +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, (combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F32))), + input_shape, data_type) +{ + InterpolationPolicy policy = InterpolationPolicy::NEAREST_NEIGHBOR; + Size2D info = Size2D(2, 2); + + // Create tensors + Tensor src = create_tensor(input_shape, data_type, 1); + Tensor dst; + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Create and configure function + NEUpsampleLayer upsample; + upsample.configure(&src, &dst, info, policy); + + // Validate valid region + const ValidRegion src_valid_region = shape_to_valid_region(src.info()->tensor_shape()); + const ValidRegion dst_valid_region = shape_to_valid_region(dst.info()->tensor_shape()); + + validate(src.info()->valid_region(), src_valid_region); + validate(dst.info()->valid_region(), dst_valid_region); +} + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(10U, 10U, 2U), 1, DataType::F32), // Mismatching data type + TensorInfo(TensorShape(10U, 10U, 2U), 1, DataType::F32), // Invalid output shape + TensorInfo(TensorShape(10U, 10U, 2U), 1, DataType::F32), // Invalid stride + TensorInfo(TensorShape(10U, 10U, 2U), 1, DataType::F32), // Invalid policy + TensorInfo(TensorShape(32U, 32U), 1, DataType::F32), + }), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(20U, 20U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(20U, 10U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(20U, 20U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(20U, 20U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(64U, 64U), 1, DataType::F32), + })), + framework::dataset::make("PadInfo", { Size2D(2, 2), + Size2D(2, 2), + Size2D(1, 1), + Size2D(2, 2), + Size2D(2, 2), + })), + framework::dataset::make("UpsamplingPolicy", { InterpolationPolicy::NEAREST_NEIGHBOR, + InterpolationPolicy::NEAREST_NEIGHBOR, + InterpolationPolicy::NEAREST_NEIGHBOR, + InterpolationPolicy::BILINEAR, + InterpolationPolicy::NEAREST_NEIGHBOR, + })), + framework::dataset::make("Expected", { false, false, false, false, true })), + input_info, output_info, pad_info, policy, expected) +{ + bool is_valid = bool(NEUpsampleLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), pad_info, policy)); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +template +using NEUpsampleLayerFixture = UpsampleLayerFixture; + +TEST_SUITE(Float) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, NEUpsampleLayerFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + framework::dataset::make("PadInfo", { Size2D(2, 2) })), + framework::dataset::make("UpsamplingPolicy", { InterpolationPolicy::NEAREST_NEIGHBOR }))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() // FP32 + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEUpsampleLayerFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataType", + DataType::F16)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + framework::dataset::make("PadInfo", { Size2D(2, 2) })), + framework::dataset::make("UpsamplingPolicy", { InterpolationPolicy::NEAREST_NEIGHBOR }))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() // FP16 +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ +TEST_SUITE_END() // Float + +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEUpsampleLayerFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + framework::dataset::make("PadInfo", { Size2D(2, 2) })), + framework::dataset::make("UpsamplingPolicy", { InterpolationPolicy::NEAREST_NEIGHBOR }))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() // Quantized +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE_END() // UpsampleLayer +TEST_SUITE_END() // NEON + +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/fixtures/UpsampleLayerFixture.h b/tests/validation/fixtures/UpsampleLayerFixture.h index 8fc3565e18..40229e2866 100644 --- a/tests/validation/fixtures/UpsampleLayerFixture.h +++ b/tests/validation/fixtures/UpsampleLayerFixture.h @@ -45,12 +45,12 @@ class UpsampleLayerFixture : public framework::Fixture public: template void setup(TensorShape input_shape, DataType data_type, DataLayout data_layout, - Size2D info, const InterpolationPolicy &upsampling_policy) + Size2D info, const InterpolationPolicy &policy) { _data_type = data_type; - _target = compute_target(input_shape, info, upsampling_policy, data_type, data_layout); - _reference = compute_reference(input_shape, info, upsampling_policy, data_type); + _target = compute_target(input_shape, info, policy, data_type, data_layout); + _reference = compute_reference(input_shape, info, policy, data_type); } protected: @@ -61,7 +61,7 @@ protected: } TensorType compute_target(TensorShape input_shape, - const Size2D &info, const InterpolationPolicy &upsampling_policy, DataType data_type, DataLayout data_layout) + const Size2D &info, const InterpolationPolicy &policy, DataType data_type, DataLayout data_layout) { if(data_layout == DataLayout::NHWC) { @@ -74,7 +74,7 @@ protected: // Create and configure function FunctionType upsample; - upsample.configure(&src, &dst, info, upsampling_policy); + upsample.configure(&src, &dst, info, policy); ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); @@ -96,7 +96,7 @@ protected: } SimpleTensor compute_reference(const TensorShape &input_shape, - const Size2D &info, const InterpolationPolicy &upsampling_policy, DataType data_type) + const Size2D &info, const InterpolationPolicy &policy, DataType data_type) { // Create reference SimpleTensor src{ input_shape, data_type }; @@ -104,7 +104,7 @@ protected: // Fill reference fill(src, 0); - return reference::upsample_layer(src, info, upsampling_policy); + return reference::upsample_layer(src, info, policy); } TensorType _target{}; diff --git a/tests/validation/reference/UpsampleLayer.cpp b/tests/validation/reference/UpsampleLayer.cpp index 3a340d0905..876f6d794a 100644 --- a/tests/validation/reference/UpsampleLayer.cpp +++ b/tests/validation/reference/UpsampleLayer.cpp @@ -35,10 +35,10 @@ namespace reference { template SimpleTensor upsample_layer(const SimpleTensor &src, - const Size2D &info, const InterpolationPolicy upsampling_policy) + const Size2D &info, const InterpolationPolicy policy) { - ARM_COMPUTE_ERROR_ON(upsampling_policy != InterpolationPolicy::NEAREST_NEIGHBOR); - ARM_COMPUTE_UNUSED(upsampling_policy); + ARM_COMPUTE_ERROR_ON(policy != InterpolationPolicy::NEAREST_NEIGHBOR); + ARM_COMPUTE_UNUSED(policy); TensorShape output_shape = src.shape(); output_shape.set(0, src.shape().x() * info.x()); @@ -77,9 +77,11 @@ SimpleTensor upsample_layer(const SimpleTensor &src, } template SimpleTensor upsample_layer(const SimpleTensor &src, - const Size2D &info, const InterpolationPolicy upsampling_policy); + const Size2D &info, const InterpolationPolicy policy); template SimpleTensor upsample_layer(const SimpleTensor &src, - const Size2D &info, const InterpolationPolicy upsampling_policy); + const Size2D &info, const InterpolationPolicy policy); +template SimpleTensor upsample_layer(const SimpleTensor &src, + const Size2D &info, const InterpolationPolicy policy); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/UpsampleLayer.h b/tests/validation/reference/UpsampleLayer.h index fc1da39186..ecb458a0c6 100644 --- a/tests/validation/reference/UpsampleLayer.h +++ b/tests/validation/reference/UpsampleLayer.h @@ -37,7 +37,7 @@ namespace reference { template SimpleTensor upsample_layer(const SimpleTensor &src, - const Size2D &info, const InterpolationPolicy upsampling_policy); + const Size2D &info, const InterpolationPolicy policy); } // namespace reference } // namespace validation } // namespace test -- cgit v1.2.1