From e07c4cbd16b25816e448c4c357b45fc3540adfe2 Mon Sep 17 00:00:00 2001 From: =?utf8?q?=EC=9E=A5=EC=A7=80=EC=84=AD/On-Device=20Lab=28SR=29/Enginee?= =?utf8?q?r/=EC=82=BC=EC=84=B1=EC=A0=84=EC=9E=90?= Date: Thu, 16 May 2019 16:41:56 +0900 Subject: [PATCH] [ACLEx] Introduce CLRNNLayerEx (#5196) This commit introduces CLRNNLayerEx fixing CLRNNLayer of v19.02. diff --git a/src/runtime/CL/functions/CLRNNLayer.cpp b/src/runtime/CL/functions/CLRNNLayer.cpp index 63f00ac8e..d89a38316 100644 --- a/src/runtime/CL/functions/CLRNNLayer.cpp +++ b/src/runtime/CL/functions/CLRNNLayer.cpp @@ -44,8 +44,8 @@ CLRNNLayer::CLRNNLayer(std::shared_ptr memory_manager) Status CLRNNLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *recurrent_weights, const ITensorInfo *bias, const ITensorInfo *hidden_state, const ITensorInfo *output, const ActivationLayerInfo &info) { - const int idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); - const int idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); + const int idx_width = 0; + const int idx_height = 1; ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, recurrent_weights, bias, hidden_state, output); ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != weights->dimension(idx_width)); ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_height) != recurrent_weights->dimension(idx_width)); @@ -72,7 +72,7 @@ void CLRNNLayer::configure(const ICLTensor *input, const ICLTensor *weights, con ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, recurrent_weights, bias, hidden_state, output); ARM_COMPUTE_ERROR_THROW_ON(CLRNNLayer::validate(input->info(), weights->info(), recurrent_weights->info(), bias->info(), hidden_state->info(), output->info(), info)); - const int idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT); + const int idx_height = 1; TensorShape shape = compute_rnn_shape(recurrent_weights->info(), hidden_state->info()->dimension(idx_height)); _is_prepared = false; Signed-off-by: jiseob.jang --- .../runtime/CL/functions/CLRNNLayerEx.h | 104 +++++++++++++++ .../src/runtime/CL/functions/CLRNNLayerEx.cpp | 147 +++++++++++++++++++++ 2 files changed, 251 insertions(+) create mode 100644 libs/ARMComputeEx/arm_compute/runtime/CL/functions/CLRNNLayerEx.h create mode 100644 libs/ARMComputeEx/src/runtime/CL/functions/CLRNNLayerEx.cpp diff --git a/libs/ARMComputeEx/arm_compute/runtime/CL/functions/CLRNNLayerEx.h b/libs/ARMComputeEx/arm_compute/runtime/CL/functions/CLRNNLayerEx.h new file mode 100644 index 0000000..d9857e6 --- /dev/null +++ b/libs/ARMComputeEx/arm_compute/runtime/CL/functions/CLRNNLayerEx.h @@ -0,0 +1,104 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * 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_CLRNN_LAYER_EX_H__ +#define __ARM_COMPUTE_CLRNN_LAYER_EX_H__ + +#include "arm_compute/core/CL/kernels/CLActivationLayerKernel.h" +#include "arm_compute/core/CL/kernels/CLCopyKernel.h" +#include "arm_compute/core/CL/kernels/CLElementwiseOperationKernel.h" +#include "arm_compute/runtime/CL/ICLSimpleFunction.h" +#include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h" +#include "arm_compute/runtime/CL/functions/CLGEMM.h" + +namespace arm_compute +{ +class ICLTensor; + +/** Basic function to run @ref CLRNNLayerEx */ +class CLRNNLayerEx : public IFunction +{ +public: + /** Default constructor */ + CLRNNLayerEx(std::shared_ptr memory_manager = nullptr); + /** Initialize the function + * + * @param[in] input Input is a 2-D tensor of shape [input_size, batch_size]. Data + * types supported: F16/F32 + * @param[in] weights Weights tensor of shape [input_size, num_units] that + * multiplies the input. Data types supported: Same as @p input + * @param[in] recurrent_weights Weights tensor of shape [num_units, num_units] that multiplies + * the current 'state'. Data types supported: Same as @p input + * @param[in] bias Bias vector of shape [num_units]. Data types supported: Same + * as @p input + * @param[out] output Output tensor of shape [num_units, batch_size]. Data types + * supported: Same as @p input + * @param[in,out] hidden_state Output tensor of shape [num_units, batch_size]. Data types + * supported: Same as @p input + * @param[in] info Activation layer parameter. + */ + void configure(const ICLTensor *input, const ICLTensor *weights, + const ICLTensor *recurrent_weights, const ICLTensor *bias, ICLTensor *hidden_state, + ICLTensor *output, ActivationLayerInfo &info); + /** Initialize the function + * + * @param[in] input Input is a 2-D tensor of shape [input_size, batch_size]. Data + * types supported: F16/F32 + * @param[in] weights Weights tensor of shape [input_size, num_units] that multiplies + * the input. Data types supported: Same as @p input + * @param[in] recurrent_weights Weights tensor of shape [num_units, num_units] that multiplies the + * current 'state'. Data types supported: Same as @p input + * @param[in] bias Bias vector of shape [num_units]. Data types supported: Same as @p + * input + * @param[in] output Output tensor of shape [num_units, batch_size]. Data types + * supported: Same as @p input + * @param[in] hidden_state Output tensor of shape [num_units, batch_size]. Data types + * supported: Same as @p input + * @param[in] info Activation layer parameter. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *weights, + const ITensorInfo *recurrent_weights, const ITensorInfo *bias, + const ITensorInfo *hidden_state, const ITensorInfo *output, + const ActivationLayerInfo &info); + + // Inherited methods overridden: + void run() override; + void prepare() override; + +private: + CLMemoryGroup _memory_group; + CLGEMM _gemm_state_f; + CLSaturatedArithmeticOperationKernel _add_kernel; + CLActivationLayerKernel _activation_kernel; + CLFullyConnectedLayer _fully_connected_kernel; + CLCopyKernel _copy_kernel; + CLTensor _fully_connected_out; + CLTensor _gemm_output; + CLTensor _add_output; + bool _is_prepared; +}; +} +#endif /* __ARM_COMPUTE_CLRNN_LAYER_EX_H__ */ diff --git a/libs/ARMComputeEx/src/runtime/CL/functions/CLRNNLayerEx.cpp b/libs/ARMComputeEx/src/runtime/CL/functions/CLRNNLayerEx.cpp new file mode 100644 index 0000000..de9cdde --- /dev/null +++ b/libs/ARMComputeEx/src/runtime/CL/functions/CLRNNLayerEx.cpp @@ -0,0 +1,147 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * 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/CL/functions/CLRNNLayerEx.h" + +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/runtime/CL/CLScheduler.h" +#include "support/ToolchainSupport.h" + +#include + +using namespace arm_compute; +using namespace arm_compute::misc::shape_calculator; + +CLRNNLayerEx::CLRNNLayerEx(std::shared_ptr memory_manager) + : _memory_group(std::move(memory_manager)), _gemm_state_f(), _add_kernel(), + _activation_kernel(), _fully_connected_kernel(), _copy_kernel(), _fully_connected_out(), + _gemm_output(), _add_output(), _is_prepared(false) +{ +} + +Status CLRNNLayerEx::validate(const ITensorInfo *input, const ITensorInfo *weights, + const ITensorInfo *recurrent_weights, const ITensorInfo *bias, + const ITensorInfo *hidden_state, const ITensorInfo *output, + const ActivationLayerInfo &info) +{ + const int idx_width = 0; + const int idx_height = 1; + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, recurrent_weights, bias, hidden_state, + output); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != weights->dimension(idx_width)); + ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_height) != + recurrent_weights->dimension(idx_width)); + ARM_COMPUTE_RETURN_ERROR_ON(recurrent_weights->dimension(idx_width) != + recurrent_weights->dimension(1)); + ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() != 1); + ARM_COMPUTE_RETURN_ERROR_ON(bias->dimension(idx_width) != weights->dimension(idx_height)); + ARM_COMPUTE_RETURN_ERROR_ON(hidden_state->dimension(idx_width) != weights->dimension(idx_height)); + ARM_COMPUTE_RETURN_ERROR_ON(hidden_state->dimension(idx_height) != input->dimension(idx_height)); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), + hidden_state->tensor_shape()); + + auto shape_info = + TensorInfo(compute_rnn_shape(recurrent_weights, hidden_state->dimension(idx_height)), 1, + input->data_type()); + + ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, weights, bias, &shape_info)); + ARM_COMPUTE_RETURN_ON_ERROR( + CLGEMM::validate(hidden_state, recurrent_weights, nullptr, &shape_info, 1.f, 0.f)); + ARM_COMPUTE_RETURN_ON_ERROR(CLSaturatedArithmeticOperationKernel::validate( + ArithmeticOperation::ADD, &shape_info, &shape_info, &shape_info, ConvertPolicy::SATURATE)); + ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&shape_info, &shape_info, info)); + + return Status{}; +} + +void CLRNNLayerEx::configure(const ICLTensor *input, const ICLTensor *weights, + const ICLTensor *recurrent_weights, const ICLTensor *bias, + ICLTensor *hidden_state, ICLTensor *output, ActivationLayerInfo &info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, recurrent_weights, bias, hidden_state, output); + ARM_COMPUTE_ERROR_THROW_ON(CLRNNLayerEx::validate(input->info(), weights->info(), + recurrent_weights->info(), bias->info(), + hidden_state->info(), output->info(), info)); + + const int idx_height = 1; + TensorShape shape = + compute_rnn_shape(recurrent_weights->info(), hidden_state->info()->dimension(idx_height)); + + _is_prepared = false; + + _fully_connected_out.allocator()->init(TensorInfo(shape, 1, input->info()->data_type())); + _gemm_output.allocator()->init(TensorInfo(shape, 1, input->info()->data_type())); + + // Manage intermediate buffers and configure + _memory_group.manage(&_fully_connected_out); + _fully_connected_kernel.configure(input, weights, bias, &_fully_connected_out); + + _memory_group.manage(&_gemm_output); + _gemm_state_f.configure(hidden_state, recurrent_weights, nullptr, &_gemm_output, 1.f, 0.f); + + _add_output.allocator()->init(TensorInfo(shape, 1, input->info()->data_type())); + _memory_group.manage(&_add_output); + + _add_kernel.configure(ArithmeticOperation::ADD, &_fully_connected_out, &_gemm_output, + &_add_output, ConvertPolicy::SATURATE); + + _fully_connected_out.allocator()->allocate(); + _gemm_output.allocator()->allocate(); + + _activation_kernel.configure(&_add_output, hidden_state, info); + _add_output.allocator()->allocate(); + + _copy_kernel.configure(hidden_state, output); +} + +void CLRNNLayerEx::run() +{ + prepare(); + + _memory_group.acquire(); + + _fully_connected_kernel.run(); + _gemm_state_f.run(); + CLScheduler::get().enqueue(_add_kernel); + CLScheduler::get().enqueue(_activation_kernel); + + // copy hidden out to output + CLScheduler::get().enqueue(_copy_kernel); + + _memory_group.release(); +} + +void CLRNNLayerEx::prepare() +{ + if (!_is_prepared) + { + _fully_connected_kernel.prepare(); + _gemm_state_f.prepare(); + + _is_prepared = true; + } +} -- 2.7.4