/* * Copyright (c) 2018-2021 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/CLLSTMLayer.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/utils/misc/InfoHelpers.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/core/utils/quantization/AsymmHelpers.h" #include "arm_compute/runtime/CL/CLScheduler.h" #include "src/core/CL/kernels/CLFillBorderKernel.h" #include "src/gpu/cl/kernels/ClTransposeKernel.h" #include "src/common/utils/Log.h" namespace arm_compute { using namespace arm_compute::misc::shape_calculator; using namespace arm_compute::utils::info_helpers; CLLSTMLayer::CLLSTMLayer(std::shared_ptr memory_manager) : _memory_group(std::move(memory_manager)), _fully_connected_input_gate(), _accum_input_gate1(), _subtract_input_gate(), _pixelwise_mul_input_gate(), _activation_input_gate(), _fully_connected_forget_gate(), _accum_forget_gate1(), _pixelwise_mul_forget_gate(), _activation_forget_gate(), _fully_connected_cell_state(), _gemm_cell_state1(), _transpose_cell_state(std::make_unique()), _accum_cell_state1(), _accum_cell_state2(), _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(), _pixelwise_mul_cell_state2(), _fully_connected_output(), _pixelwise_mul_output_state1(), _accum_output1(), _activation_output(), _activation_output_state(), _pixelwise_mul_output_state2(), _fully_connected_output_state(), _projection_clip(), _copy_cell_state(), _copy_output(), _concat_scratch_buffer(), _concat_inputs_forget_gate(), _concat_weights_forget_gate(), _concat_weights_input_gate(), _concat_weights_output(), _ones_fill(), _mean_std_norm_input_gate(), _pixelwise_mul_input_gate_coeff(), _accum_input_gate_bias(), _mean_std_norm_forget_gate(), _pixelwise_mul_forget_gate_coeff(), _accum_forget_gate_bias(), _mean_std_norm_cell_gate(), _pixelwise_mul_cell_gate_coeff(), _accum_cell_gate_bias(), _mean_std_norm_output_gate(), _pixelwise_mul_output_gate_coeff(), _accum_output_gate_bias(), _input_gate_out1(), _input_gate_out2(), _input_gate_out3(), _input_gate_out4(), _forget_gate_out1(), _forget_gate_out2(), _forget_gate_out3(), _forget_gate_out4(), _forget_gate_out5(), _forget_gate_out6(), _cell_state_out1(), _cell_state_out2(), _cell_state_out3(), _cell_state_out4(), _cell_state_out5(), _output1(), _output2(), _output3(), _output4(), _cell_state_activation(), _output_state1(), _ones(), _input_layer_norm_out1(), _input_layer_norm_out2(), _forget_layer_norm_out1(), _forget_layer_norm_out2(), _cell_layer_norm_out1(), _cell_layer_norm_out2(), _output_layer_norm_out1(), _output_layer_norm_out2(), _run_peephole_opt(false), _run_cifg_opt(false), _perform_cell_clipping(false), _has_projection_weights(false), _perform_projection_clipping(false), _is_prepared(false), _is_layer_norm_lstm(false) { } CLLSTMLayer::~CLLSTMLayer() = default; void CLLSTMLayer::configure(const ICLTensor *input, const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, const ICLTensor *output_state_in, ICLTensor *cell_state_in, ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output, const LSTMParams &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold) { configure(CLKernelLibrary::get().get_compile_context(), input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights, forget_gate_bias, cell_bias, output_gate_bias, output_state_in, cell_state_in, scratch_buffer, output_state_out, cell_state_out, output, lstm_params, activation_info, cell_threshold, projection_threshold); } void CLLSTMLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, const ICLTensor *output_state_in, ICLTensor *cell_state_in, ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output, const LSTMParams &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights, forget_gate_bias, cell_bias, output_gate_bias, output_state_in, cell_state_in, scratch_buffer, output_state_out, cell_state_out, output); ARM_COMPUTE_LOG_PARAMS(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights, forget_gate_bias, cell_bias, output_gate_bias, output_state_in, cell_state_in, scratch_buffer, output_state_out, cell_state_out, output, lstm_params, activation_info, cell_threshold, projection_threshold); _is_layer_norm_lstm = lstm_params.use_layer_norm(); // Set lstm parameters LSTMParams lstm_params_info{}; build_lstm_params_tensor_info(lstm_params, &lstm_params_info); // Validate ARM_COMPUTE_ERROR_THROW_ON(CLLSTMLayer::validate(input->info(), input_to_forget_weights->info(), input_to_cell_weights->info(), input_to_output_weights->info(), recurrent_to_forget_weights->info(), recurrent_to_cell_weights->info(), recurrent_to_output_weights->info(), forget_gate_bias->info(), cell_bias->info(), output_gate_bias->info(), output_state_in->info(), cell_state_in->info(), scratch_buffer->info(), output_state_out->info(), cell_state_out->info(), output->info(), lstm_params_info, activation_info, cell_threshold, projection_threshold)); const TensorShape cell_state_shape = cell_state_in->info()->tensor_shape(); // Configure block that calculates the forget gate // forget_gate = Activation(input * input_to_forget_weights + output_state_in * recurrent_to_forget_weights + PixelWiseMul(cell_state, cell_to_forget_weights) + forget_gate_bias) // We optimize this as follows: // forget_gate = Activation( (input,output_state_in) * (input_to_forget_weights,recurrent_to_forget_weights) + PixelWiseMul(cell_state, cell_to_forget_weights) + forget_gate_bias _forget_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _forget_gate_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _forget_gate_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); std::vector inputs_vector; inputs_vector.emplace_back(input); inputs_vector.emplace_back(output_state_in); const TensorShape concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, 0); _forget_gate_out2.allocator()->init(TensorInfo(concat_shape, 1, input->info()->data_type())); _memory_group.manage(&_forget_gate_out2); _concat_inputs_forget_gate.configure(compile_context, inputs_vector, &_forget_gate_out2, Window::DimX); std::vector weights_vector; weights_vector.emplace_back(input_to_forget_weights); weights_vector.emplace_back(recurrent_to_forget_weights); const TensorShape weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(weights_vector, 0); _forget_gate_out6.allocator()->init(TensorInfo(weights_concat_shape, 1, input->info()->data_type())); _concat_weights_forget_gate.configure(compile_context, weights_vector, &_forget_gate_out6, Window::DimX); _memory_group.manage(&_forget_gate_out5); _fully_connected_forget_gate.configure(compile_context, &_forget_gate_out2, &_forget_gate_out6, (_is_layer_norm_lstm) ? nullptr : forget_gate_bias, &_forget_gate_out5); _memory_group.manage(&_forget_gate_out1); _memory_group.manage(&_forget_gate_out3); _forget_gate_out6.allocator()->allocate(); CLTensor *forget_gate_out = &_forget_gate_out5; if(lstm_params.has_peephole_opt()) { _forget_gate_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _run_peephole_opt = true; _memory_group.manage(&_forget_gate_out4); _pixelwise_mul_forget_gate.configure(compile_context, cell_state_in, lstm_params.cell_to_forget_weights(), &_forget_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); _accum_forget_gate1.configure(compile_context, &_forget_gate_out5, &_forget_gate_out4, &_forget_gate_out3, ConvertPolicy::SATURATE); _forget_gate_out4.allocator()->allocate(); _forget_gate_out5.allocator()->allocate(); forget_gate_out = &_forget_gate_out3; } else { _forget_gate_out3.allocator()->allocate(); } if(_is_layer_norm_lstm) { _forget_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _forget_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _memory_group.manage(&_forget_layer_norm_out1); _memory_group.manage(&_forget_layer_norm_out2); _mean_std_norm_forget_gate.configure(compile_context, forget_gate_out); _pixelwise_mul_forget_gate_coeff.configure(compile_context, forget_gate_out, lstm_params.forget_layer_norm_weights(), &_forget_layer_norm_out1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); // forget_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before forget_gate_out->allocator()->allocate(); _accum_forget_gate_bias.configure(compile_context, &_forget_layer_norm_out1, forget_gate_bias, &_forget_layer_norm_out2, ConvertPolicy::SATURATE); _forget_layer_norm_out1.allocator()->allocate(); forget_gate_out = &_forget_layer_norm_out2; } _activation_forget_gate.configure(compile_context, forget_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); // Configure block that calculates the input gate // input_gate = Activation(input * input_to_input_weights + output_state * recurrent_to_input_weights + PixelWiseMul(cell_state, cell_to_input_weights) + input_gate_bias), without CIFG // input_gate = 1 - forget_gate, with CIFG // We optimize this as follows: // input_gate = Activation((input,output_state) * (input_to_input_weights,recurrent_to_input_weights) + PixelWiseMul(cell_state, cell_to_input_weights) + input_gate_bias), without CIFG _input_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); CLTensor *input_gate_out = &_input_gate_out1; if(lstm_params.has_cifg_opt()) { _memory_group.manage(&_input_gate_out1); _ones.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _ones_fill.configure(compile_context, &_ones, PixelValue(1, _ones.info()->data_type())); _subtract_input_gate.configure(compile_context, &_ones, forget_gate_out, &_input_gate_out1, ConvertPolicy::SATURATE); _ones.allocator()->allocate(); _run_cifg_opt = true; } else { _input_gate_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _input_gate_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); std::vector lstm_weights; lstm_weights.emplace_back(lstm_params.input_to_input_weights()); lstm_weights.emplace_back(lstm_params.recurrent_to_input_weights()); TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(lstm_weights, 0); _input_gate_out2.allocator()->init(TensorInfo(lstm_weights_concat_shape, 1, input->info()->data_type())); _concat_weights_input_gate.configure(compile_context, lstm_weights, &_input_gate_out2, Window::DimX); _memory_group.manage(&_input_gate_out1); _memory_group.manage(&_input_gate_out3); _fully_connected_input_gate.configure(compile_context, &_forget_gate_out2, &_input_gate_out2, (_is_layer_norm_lstm) ? nullptr : lstm_params.input_gate_bias(), &_input_gate_out3); _input_gate_out2.allocator()->allocate(); input_gate_out = &_input_gate_out3; if(_run_peephole_opt) { _memory_group.manage(&_input_gate_out4); _pixelwise_mul_input_gate.configure(compile_context, cell_state_in, lstm_params.cell_to_input_weights(), &_input_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); _accum_input_gate1.configure(compile_context, &_input_gate_out3, &_input_gate_out4, &_input_gate_out1, ConvertPolicy::SATURATE); _input_gate_out3.allocator()->allocate(); _input_gate_out4.allocator()->allocate(); input_gate_out = &_input_gate_out1; } else { _input_gate_out1.allocator()->allocate(); } if(_is_layer_norm_lstm) { _input_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _input_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _memory_group.manage(&_input_layer_norm_out1); _memory_group.manage(&_input_layer_norm_out2); _mean_std_norm_input_gate.configure(compile_context, input_gate_out); _pixelwise_mul_input_gate_coeff.configure(compile_context, input_gate_out, lstm_params.input_layer_norm_weights(), &_input_layer_norm_out1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); // input_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before input_gate_out->allocator()->allocate(); _accum_input_gate_bias.configure(compile_context, &_input_layer_norm_out1, lstm_params.input_gate_bias(), &_input_layer_norm_out2, ConvertPolicy::SATURATE); _input_layer_norm_out1.allocator()->allocate(); input_gate_out = &_input_layer_norm_out2; } _activation_input_gate.configure(compile_context, input_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); } // Configure block that calculates the cell state // cell_state = Clip((PixelwiseMul(input_gate, Activation(input * input_to_cell_weights + output_state_in * recurrent_to_cell_weights + cell_bias)) + PixelwiseMul(forget_gate, cell_state)), cell_threshold) TensorShape cell_state1_shape = compute_transposed_shape(*recurrent_to_output_weights->info()); _cell_state_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _cell_state_out2.allocator()->init(TensorInfo(cell_state1_shape, 1, input->info()->data_type())); _cell_state_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _cell_state_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _cell_state_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _memory_group.manage(&_cell_state_out1); _fully_connected_cell_state.configure(compile_context, input, input_to_cell_weights, (_is_layer_norm_lstm) ? nullptr : cell_bias, &_cell_state_out1); _memory_group.manage(&_cell_state_out2); _transpose_cell_state->configure(compile_context, recurrent_to_cell_weights->info(), _cell_state_out2.info()); _recurrent_to_cell_weights = recurrent_to_cell_weights; _memory_group.manage(&_cell_state_out3); _gemm_cell_state1.configure(compile_context, output_state_in, &_cell_state_out2, nullptr, &_cell_state_out3, 1.f, 0.f); _cell_state_out2.allocator()->allocate(); _memory_group.manage(&_cell_state_out4); _accum_cell_state1.configure(compile_context, &_cell_state_out1, &_cell_state_out3, &_cell_state_out4, ConvertPolicy::SATURATE); CLTensor *cell_state_out_ptr = &_cell_state_out4; if(_is_layer_norm_lstm) { _cell_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _cell_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _memory_group.manage(&_cell_layer_norm_out1); _memory_group.manage(&_cell_layer_norm_out2); _mean_std_norm_cell_gate.configure(compile_context, cell_state_out_ptr); _pixelwise_mul_cell_gate_coeff.configure(compile_context, cell_state_out_ptr, lstm_params.cell_layer_norm_weights(), &_cell_layer_norm_out1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); // cell_state_out_ptr is going to be reassigned, so allocate the tensor that it was assigned to before cell_state_out_ptr->allocator()->allocate(); _accum_cell_gate_bias.configure(compile_context, &_cell_layer_norm_out1, cell_bias, &_cell_layer_norm_out2, ConvertPolicy::SATURATE); _cell_layer_norm_out1.allocator()->allocate(); cell_state_out_ptr = &_cell_layer_norm_out2; } _activation_cell_state.configure(compile_context, cell_state_out_ptr, nullptr, activation_info); _memory_group.manage(&_cell_state_out5); _pixelwise_mul_cell_state1.configure(compile_context, cell_state_out_ptr, input_gate_out, &_cell_state_out5, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); cell_state_out_ptr->allocator()->allocate(); _pixelwise_mul_cell_state2.configure(compile_context, forget_gate_out, cell_state_in, &_cell_state_out3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); _accum_cell_state2.configure(compile_context, &_cell_state_out5, &_cell_state_out3, &_cell_state_out1, ConvertPolicy::SATURATE); _cell_state_out3.allocator()->allocate(); _cell_state_out5.allocator()->allocate(); // Perform clipping if(cell_threshold != 0.f) { _perform_cell_clipping = true; _cell_clip.configure(compile_context, &_cell_state_out1, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, cell_threshold, -cell_threshold)); } // Configure block that calculates the output // output_state_out = Activation(input * input_to_output_weights + output_state_in * recurrent_to_output_weights + PixelWiseMul(cell_state, cell_to_output_weights) + output_gate_bias) // We optimize this as follows: // output_state_out = Activation( (input,output_state_in) * (input_to_output_weights, recurrent_to_output_weights) + PixelWiseMul(cell_state, cell_to_output_weights) + output_gate_bias) _output1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _output4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); std::vector in_out_weights; in_out_weights.emplace_back(input_to_output_weights); in_out_weights.emplace_back(recurrent_to_output_weights); TensorShape in_out_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(in_out_weights, 0); _output2.allocator()->init(TensorInfo(in_out_weights_concat_shape, 1, input->info()->data_type())); _concat_weights_output.configure(compile_context, in_out_weights, &_output2, Window::DimX); _memory_group.manage(&_output1); _memory_group.manage(&_output4); _fully_connected_output.configure(compile_context, &_forget_gate_out2, &_output2, (_is_layer_norm_lstm) ? nullptr : output_gate_bias, &_output4); _output2.allocator()->allocate(); _forget_gate_out2.allocator()->allocate(); CLTensor *output_gate_out = &_output4; if(lstm_params.has_peephole_opt()) { _output3.allocator()->init(TensorInfo(_cell_state_out1.info()->tensor_shape(), 1, input->info()->data_type())); _memory_group.manage(&_output3); _pixelwise_mul_output_state1.configure(compile_context, &_cell_state_out1, lstm_params.cell_to_output_weights(), &_output3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); _accum_output1.configure(compile_context, &_output4, &_output3, &_output1, ConvertPolicy::SATURATE); _output4.allocator()->allocate(); output_gate_out = &_output1; // Allocate intermediate buffers _output3.allocator()->allocate(); } else { _output1.allocator()->allocate(); } if(_is_layer_norm_lstm) { _output_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _output_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _memory_group.manage(&_output_layer_norm_out1); _memory_group.manage(&_output_layer_norm_out2); _mean_std_norm_output_gate.configure(compile_context, output_gate_out); _pixelwise_mul_output_gate_coeff.configure(compile_context, output_gate_out, lstm_params.output_layer_norm_weights(), &_output_layer_norm_out1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); // output_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before output_gate_out->allocator()->allocate(); _accum_output_gate_bias.configure(compile_context, &_output_layer_norm_out1, output_gate_bias, &_output_layer_norm_out2, ConvertPolicy::SATURATE); _output_layer_norm_out1.allocator()->allocate(); output_gate_out = &_output_layer_norm_out2; } _activation_output.configure(compile_context, output_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); // Configure block that calculates the output state /** lstm_res = PixelwiseMul(output, Activation(cell_state)) * * -- Clip(lstm_res * projection_weights + projection_bias, projection_threshold) , if there is a projection * / * output_state = -- * \ * -- lstm_res , otherwise */ ICLTensor *output_state_out_tmp = lstm_params.has_projection() ? &_output_state1 : output_state_out; _cell_state_activation.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _output_state1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); _memory_group.manage(&_cell_state_activation); _activation_output_state.configure(compile_context, &_cell_state_out1, &_cell_state_activation, activation_info); _pixelwise_mul_output_state2.configure(compile_context, &_cell_state_activation, output_gate_out, output_state_out_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN); _cell_state_activation.allocator()->allocate(); if(lstm_params.has_projection()) { _has_projection_weights = true; _fully_connected_output_state.configure(compile_context, output_state_out_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out); _output_state1.allocator()->allocate(); // Perform clipping if(projection_threshold != 0.f) { _perform_projection_clipping = true; _projection_clip.configure(compile_context, output_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold)); } } // Copy cell state and output _copy_cell_state.configure(compile_context, &_cell_state_out1, cell_state_out); _copy_output.configure(compile_context, output_state_out, output); // Vector for holding the tensors to store in scratch buffer std::vector scratch_inputs; if(!lstm_params.has_cifg_opt()) { scratch_inputs.emplace_back(input_gate_out); } scratch_inputs.emplace_back(&_cell_state_out1); scratch_inputs.emplace_back(forget_gate_out); scratch_inputs.emplace_back(output_gate_out); _concat_scratch_buffer.configure(compile_context, scratch_inputs, scratch_buffer, Window::DimX); input_gate_out->allocator()->allocate(); _cell_state_out1.allocator()->allocate(); forget_gate_out->allocator()->allocate(); output_gate_out->allocator()->allocate(); } Status CLLSTMLayer::validate(const ITensorInfo *input, const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights, const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights, const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias, const ITensorInfo *output_state_in, const ITensorInfo *cell_state_in, const ITensorInfo *scratch_buffer, const ITensorInfo *output_state_out, const ITensorInfo *cell_state_out, const ITensorInfo *output, const LSTMParams &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights, forget_gate_bias, cell_bias, output_gate_bias, output_state_in, cell_state_in, scratch_buffer, output_state_out, cell_state_out, output); // Check data types ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights, forget_gate_bias, cell_bias, output_gate_bias, output_state_in, cell_state_in, scratch_buffer, output_state_out, cell_state_out, output); // Check dimensions ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(input_to_forget_weights->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(input_to_cell_weights->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(input_to_output_weights->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_forget_weights->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_cell_weights->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_output_weights->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(forget_gate_bias->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(cell_bias->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(output_gate_bias->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(cell_state_in->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(scratch_buffer->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(output_state_out->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(cell_state_out->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(cell_bias->dimension(0) * 4 != scratch_buffer->dimension(0) && cell_bias->dimension(0) * 3 != scratch_buffer->dimension(0)); const unsigned int num_batches = input->dimension(1); const unsigned int num_cells = input_to_output_weights->dimension(1); if(lstm_params.use_layer_norm()) { // If CIFG is used, input layer normalization weights tensor is omitted if(lstm_params.has_cifg_opt()) { ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights() != nullptr); } else { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.input_layer_norm_weights()); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights()->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights()->dimension(0) != num_cells); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, lstm_params.input_layer_norm_weights()); } ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.forget_layer_norm_weights(), lstm_params.cell_layer_norm_weights(), lstm_params.output_layer_norm_weights()); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, lstm_params.forget_layer_norm_weights(), lstm_params.cell_layer_norm_weights(), lstm_params.output_layer_norm_weights()); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.forget_layer_norm_weights()->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_layer_norm_weights()->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.output_layer_norm_weights()->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.forget_layer_norm_weights()->dimension(0) != num_cells); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_layer_norm_weights()->dimension(0) != num_cells); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.output_layer_norm_weights()->dimension(0) != num_cells); } // Check peephole optimization if(lstm_params.has_peephole_opt()) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.cell_to_output_weights(), lstm_params.cell_to_forget_weights()); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_forget_weights()->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_output_weights()->num_dimensions() > 1); } TensorShape units_out_transposed_shape = compute_transposed_shape(*recurrent_to_output_weights); TensorShape num_units_transposed_shape = compute_transposed_shape(*forget_gate_bias); const TensorInfo units_out_transposed_info = TensorInfo(units_out_transposed_shape, 1, input->data_type()); const TensorInfo num_units_transposed_info = TensorInfo(num_units_transposed_shape, 1, input->data_type()); TensorInfo input_gate = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type()); TensorInfo forget_gate = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type()); TensorInfo output_gate_tmp = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type()); TensorInfo cell_state_tmp = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type()); // Validate forget gate ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_forget_weights, (lstm_params.use_layer_norm()) ? nullptr : forget_gate_bias, &forget_gate)); std::vector inputs_vector; inputs_vector.emplace_back(input); inputs_vector.emplace_back(output_state_in); const TensorShape concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, 0); TensorInfo forget_gate_concat = TensorInfo(concat_shape, 1, input->data_type()); ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(inputs_vector, &forget_gate_concat, Window::DimX)); if(lstm_params.has_peephole_opt()) { ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &forget_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE)); } if(lstm_params.use_layer_norm()) { ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&forget_gate)); ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&forget_gate, lstm_params.forget_layer_norm_weights(), &forget_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, forget_gate_bias, &forget_gate, ConvertPolicy::SATURATE)); } ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&forget_gate, &forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); // Validate input gate if(!lstm_params.has_cifg_opt()) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.input_to_input_weights(), lstm_params.recurrent_to_input_weights(), lstm_params.input_gate_bias()); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_to_input_weights()->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.recurrent_to_input_weights()->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_gate_bias()->num_dimensions() > 1); std::vector lstm_weights; lstm_weights.emplace_back(lstm_params.input_to_input_weights()); lstm_weights.emplace_back(lstm_params.recurrent_to_input_weights()); TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(lstm_weights, 0); TensorInfo lstm_gate_concat = TensorInfo(lstm_weights_concat_shape, 1, input->data_type()); ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(lstm_weights, &lstm_gate_concat, Window::DimX)); ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, lstm_params.input_to_input_weights(), (lstm_params.use_layer_norm()) ? nullptr : lstm_params.input_gate_bias(), &input_gate)); if(lstm_params.has_peephole_opt()) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.cell_to_input_weights()); ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_input_weights()->num_dimensions() > 1); ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_input_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, &input_gate, &input_gate, ConvertPolicy::SATURATE)); } if(lstm_params.use_layer_norm()) { ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&input_gate)); ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&input_gate, lstm_params.input_layer_norm_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, lstm_params.input_gate_bias(), &input_gate, ConvertPolicy::SATURATE)); } ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&input_gate, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); } else { ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticSubtraction::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE)); } // Validate cell state ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_cell_weights, (lstm_params.use_layer_norm()) ? nullptr : cell_bias, &cell_state_tmp)); ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(output_state_in, &units_out_transposed_info, nullptr, &cell_state_tmp, 1.f, 0.f, GEMMInfo())); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE)); if(lstm_params.use_layer_norm()) { ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&cell_state_tmp)); ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, lstm_params.cell_layer_norm_weights(), &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, cell_bias, &cell_state_tmp, ConvertPolicy::SATURATE)); } ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, nullptr, activation_info)); ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, &input_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, &forget_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE)); if(cell_threshold != 0.f) { ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, cell_threshold, -cell_threshold))); } std::vector in_out_weights; in_out_weights.emplace_back(input_to_output_weights); in_out_weights.emplace_back(recurrent_to_output_weights); TensorShape in_out_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(in_out_weights, 0); TensorInfo in_out_gate_concat = TensorInfo(in_out_weights_concat_shape, 1, input->data_type()); ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(in_out_weights, &in_out_gate_concat, Window::DimX)); // Validate output gate tmp ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_output_weights, (lstm_params.use_layer_norm()) ? nullptr : output_gate_bias, &output_gate_tmp)); if(lstm_params.has_peephole_opt()) { ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, lstm_params.cell_to_output_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, &output_gate_tmp, &output_gate_tmp, ConvertPolicy::SATURATE)); } if(lstm_params.use_layer_norm()) { ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&output_gate_tmp)); ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&output_gate_tmp, lstm_params.output_layer_norm_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, output_gate_bias, &output_gate_tmp, ConvertPolicy::SATURATE)); } ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&output_gate_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); // Validate output state ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, &cell_state_tmp, activation_info)); ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, &output_gate_tmp, &output_gate_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN)); if(lstm_params.has_projection()) { ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(&output_gate_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out)); if(projection_threshold != 0.f) { ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output_state_out, output_state_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold))); } } // Validate copy kernel ARM_COMPUTE_RETURN_ON_ERROR(CLCopy::validate(&cell_state_tmp, cell_state_out)); ARM_COMPUTE_RETURN_ON_ERROR(CLCopy::validate(output_state_out, output)); // Validate scratch concatenation std::vector inputs_vector_info_raw; if(!lstm_params.has_cifg_opt()) { inputs_vector_info_raw.push_back(&input_gate); } inputs_vector_info_raw.push_back(&cell_state_tmp); inputs_vector_info_raw.push_back(&forget_gate); inputs_vector_info_raw.push_back(&output_gate_tmp); ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(inputs_vector_info_raw, scratch_buffer, Window::DimX)); return Status{}; } void CLLSTMLayer::run() { prepare(); MemoryGroupResourceScope scope_mg(_memory_group); _concat_inputs_forget_gate.run(); _fully_connected_forget_gate.run(); if(_run_peephole_opt) { _pixelwise_mul_forget_gate.run(); _accum_forget_gate1.run(); } if(_is_layer_norm_lstm) { _mean_std_norm_forget_gate.run(); _pixelwise_mul_forget_gate_coeff.run(); _accum_forget_gate_bias.run(); } _activation_forget_gate.run(); if(_run_cifg_opt) { _ones_fill.run(); _subtract_input_gate.run(); } else { _fully_connected_input_gate.run(); if(_run_peephole_opt) { _pixelwise_mul_input_gate.run(); _accum_input_gate1.run(); } if(_is_layer_norm_lstm) { _mean_std_norm_input_gate.run(); _pixelwise_mul_input_gate_coeff.run(); _accum_input_gate_bias.run(); } _activation_input_gate.run(); } _fully_connected_cell_state.run(); ITensorPack pack; pack.add_tensor(TensorType::ACL_SRC, _recurrent_to_cell_weights); pack.add_tensor(TensorType::ACL_DST, &_cell_state_out2); CLScheduler::get().enqueue_op(*_transpose_cell_state, pack, false); _gemm_cell_state1.run(); _accum_cell_state1.run(); if(_is_layer_norm_lstm) { _mean_std_norm_cell_gate.run(); _pixelwise_mul_cell_gate_coeff.run(); _accum_cell_gate_bias.run(); } _activation_cell_state.run(); _pixelwise_mul_cell_state1.run(); _pixelwise_mul_cell_state2.run(); _accum_cell_state2.run(); if(_perform_cell_clipping) { _cell_clip.run(); } _fully_connected_output.run(); if(_run_peephole_opt) { _pixelwise_mul_output_state1.run(); _accum_output1.run(); } if(_is_layer_norm_lstm) { _mean_std_norm_output_gate.run(); _pixelwise_mul_output_gate_coeff.run(); _accum_output_gate_bias.run(); } _activation_output.run(); _activation_output_state.run(); _pixelwise_mul_output_state2.run(); if(_has_projection_weights) { _fully_connected_output_state.run(); if(_perform_projection_clipping) { _projection_clip.run(); } } _copy_cell_state.run(); _copy_output.run(); _concat_scratch_buffer.run(); } void CLLSTMLayer::prepare() { if(!_is_prepared) { _concat_weights_forget_gate.run(); if(!_run_cifg_opt) { _concat_weights_input_gate.run(); } _concat_weights_output.run(); _is_prepared = true; } } } // namespace arm_compute