diff options
Diffstat (limited to 'src/runtime/NEON/functions/NELSTMLayer.cpp')
| -rw-r--r-- | src/runtime/NEON/functions/NELSTMLayer.cpp | 518 |
1 files changed, 342 insertions, 176 deletions
diff --git a/src/runtime/NEON/functions/NELSTMLayer.cpp b/src/runtime/NEON/functions/NELSTMLayer.cpp index d338e4fd2d..1a08cdeb06 100644 --- a/src/runtime/NEON/functions/NELSTMLayer.cpp +++ b/src/runtime/NEON/functions/NELSTMLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,20 +24,13 @@ #include "arm_compute/runtime/NEON/functions/NELSTMLayer.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/core/Validate.h" #include "arm_compute/runtime/common/LSTMParams.h" -#include "src/core/NEON/kernels/NEConvertQuantizedSignednessKernel.h" -#include "src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h" -#include "src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h" -#include "src/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h" -#include "src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h" -#include "src/core/NEON/kernels/NEGEMMLowpReductionKernel.h" -#include "src/core/NEON/kernels/NEGEMMMatrixAdditionKernel.h" -#include "src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h" -#include "src/core/NEON/kernels/NEGEMMTranspose1xWKernel.h" + +#include "src/common/utils/Log.h" namespace arm_compute { @@ -47,35 +40,122 @@ using namespace arm_compute::utils::info_helpers; NELSTMLayer::~NELSTMLayer() = default; NELSTMLayer::NELSTMLayer(std::shared_ptr<IMemoryManager> 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(), - _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(), - _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), + : _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(), + _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(), + _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) { } -void NELSTMLayer::configure(const ITensor *input, - const ITensor *input_to_forget_weights, const ITensor *input_to_cell_weights, const ITensor *input_to_output_weights, - const ITensor *recurrent_to_forget_weights, const ITensor *recurrent_to_cell_weights, const ITensor *recurrent_to_output_weights, - const ITensor *forget_gate_bias, const ITensor *cell_bias, const ITensor *output_gate_bias, - const ITensor *output_state_in, const ITensor *cell_state_in, - ITensor *scratch_buffer, ITensor *output_state_out, ITensor *cell_state_out, ITensor *output, - const LSTMParams<ITensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold) +void NELSTMLayer::configure(const ITensor *input, + const ITensor *input_to_forget_weights, + const ITensor *input_to_cell_weights, + const ITensor *input_to_output_weights, + const ITensor *recurrent_to_forget_weights, + const ITensor *recurrent_to_cell_weights, + const ITensor *recurrent_to_output_weights, + const ITensor *forget_gate_bias, + const ITensor *cell_bias, + const ITensor *output_gate_bias, + const ITensor *output_state_in, + const ITensor *cell_state_in, + ITensor *scratch_buffer, + ITensor *output_state_out, + ITensor *cell_state_out, + ITensor *output, + const LSTMParams<ITensor> &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, + 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, + 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(); @@ -84,13 +164,12 @@ void NELSTMLayer::configure(const ITensor *input, build_lstm_params_tensor_info(lstm_params, &lstm_params_info); // Validate - ARM_COMPUTE_ERROR_THROW_ON(NELSTMLayer::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)); + ARM_COMPUTE_ERROR_THROW_ON(NELSTMLayer::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(); @@ -117,20 +196,23 @@ void NELSTMLayer::configure(const ITensor *input, _concat_weights_forget_gate.configure(weights_vector, &_forget_gate_out6, Window::DimX); _memory_group.manage(&_forget_gate_out5); - _fully_connected_forget_gate.configure(&_forget_gate_out2, &_forget_gate_out6, (_is_layer_norm_lstm) ? nullptr : forget_gate_bias, &_forget_gate_out5); + _fully_connected_forget_gate.configure(&_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(); Tensor *forget_gate_out = &_forget_gate_out5; - if(lstm_params.has_peephole_opt()) + 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(cell_state_in, lstm_params.cell_to_forget_weights(), &_forget_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); - _accum_forget_gate1.configure(&_forget_gate_out5, &_forget_gate_out4, &_forget_gate_out3, ConvertPolicy::SATURATE); + _pixelwise_mul_forget_gate.configure(cell_state_in, lstm_params.cell_to_forget_weights(), &_forget_gate_out4, 1, + ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); + _accum_forget_gate1.configure(&_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; @@ -139,21 +221,25 @@ void NELSTMLayer::configure(const ITensor *input, { _forget_gate_out3.allocator()->allocate(); } - if(_is_layer_norm_lstm) + 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(forget_gate_out); - _pixelwise_mul_forget_gate_coeff.configure(forget_gate_out, lstm_params.forget_layer_norm_weights(), &_forget_layer_norm_out1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); + _pixelwise_mul_forget_gate_coeff.configure(forget_gate_out, lstm_params.forget_layer_norm_weights(), + &_forget_layer_norm_out1, 1, ConvertPolicy::SATURATE, + RoundingPolicy::TO_ZERO); // 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(&_forget_layer_norm_out1, forget_gate_bias, &_forget_layer_norm_out2, ConvertPolicy::SATURATE); + _accum_forget_gate_bias.configure(&_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(forget_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); + _activation_forget_gate.configure(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 @@ -162,7 +248,7 @@ void NELSTMLayer::configure(const ITensor *input, // 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())); Tensor *input_gate_out = &_input_gate_out1; - if(lstm_params.has_cifg_opt()) + if (lstm_params.has_cifg_opt()) { _memory_group.manage(&_input_gate_out1); _ones.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type())); @@ -184,15 +270,19 @@ void NELSTMLayer::configure(const ITensor *input, _memory_group.manage(&_input_gate_out1); _memory_group.manage(&_input_gate_out4); - _fully_connected_input_gate.configure(&_forget_gate_out2, &_input_gate_out2, (_is_layer_norm_lstm) ? nullptr : lstm_params.input_gate_bias(), &_input_gate_out3); + _fully_connected_input_gate.configure(&_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) + if (_run_peephole_opt) { _memory_group.manage(&_input_gate_out4); - _pixelwise_mul_input_gate.configure(cell_state_in, lstm_params.cell_to_input_weights(), &_input_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); - _accum_input_gate1.configure(&_input_gate_out3, &_input_gate_out4, &_input_gate_out1, ConvertPolicy::SATURATE); + _pixelwise_mul_input_gate.configure(cell_state_in, lstm_params.cell_to_input_weights(), &_input_gate_out4, + 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); + _accum_input_gate1.configure(&_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; @@ -202,21 +292,25 @@ void NELSTMLayer::configure(const ITensor *input, _input_gate_out1.allocator()->allocate(); } - if(_is_layer_norm_lstm) + 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(input_gate_out); - _pixelwise_mul_input_gate_coeff.configure(input_gate_out, lstm_params.input_layer_norm_weights(), &_input_layer_norm_out1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); + _pixelwise_mul_input_gate_coeff.configure(input_gate_out, lstm_params.input_layer_norm_weights(), + &_input_layer_norm_out1, 1, ConvertPolicy::SATURATE, + RoundingPolicy::TO_ZERO); // 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(&_input_layer_norm_out1, lstm_params.input_gate_bias(), &_input_layer_norm_out2, ConvertPolicy::SATURATE); + _accum_input_gate_bias.configure(&_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(input_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); + _activation_input_gate.configure(input_gate_out, nullptr, + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); } // Configure block that calculates the cell state @@ -229,7 +323,8 @@ void NELSTMLayer::configure(const ITensor *input, _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(input, input_to_cell_weights, (_is_layer_norm_lstm) ? nullptr : cell_bias, &_cell_state_out1); + _fully_connected_cell_state.configure(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(recurrent_to_cell_weights, &_cell_state_out2); _memory_group.manage(&_cell_state_out3); @@ -238,33 +333,40 @@ void NELSTMLayer::configure(const ITensor *input, _memory_group.manage(&_cell_state_out4); _accum_cell_state1.configure(&_cell_state_out1, &_cell_state_out3, &_cell_state_out4, ConvertPolicy::SATURATE); Tensor *cell_state_out_ptr = &_cell_state_out4; - if(_is_layer_norm_lstm) + 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(cell_state_out_ptr); - _pixelwise_mul_cell_gate_coeff.configure(cell_state_out_ptr, lstm_params.cell_layer_norm_weights(), &_cell_layer_norm_out1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); + _pixelwise_mul_cell_gate_coeff.configure(cell_state_out_ptr, lstm_params.cell_layer_norm_weights(), + &_cell_layer_norm_out1, 1, ConvertPolicy::SATURATE, + RoundingPolicy::TO_ZERO); // 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(&_cell_layer_norm_out1, cell_bias, &_cell_layer_norm_out2, ConvertPolicy::SATURATE); + _accum_cell_gate_bias.configure(&_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(cell_state_out_ptr, nullptr, activation_info); _memory_group.manage(&_cell_state_out5); - _pixelwise_mul_cell_state1.configure(cell_state_out_ptr, input_gate_out, &_cell_state_out5, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); + _pixelwise_mul_cell_state1.configure(cell_state_out_ptr, input_gate_out, &_cell_state_out5, 1, + ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); cell_state_out_ptr->allocator()->allocate(); - _pixelwise_mul_cell_state2.configure(forget_gate_out, cell_state_in, &_cell_state_out3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); + _pixelwise_mul_cell_state2.configure(forget_gate_out, cell_state_in, &_cell_state_out3, 1, ConvertPolicy::SATURATE, + RoundingPolicy::TO_ZERO); _accum_cell_state2.configure(&_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) + if (cell_threshold != 0.f) { _perform_cell_clipping = true; - _cell_clip.configure(&_cell_state_out1, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold, cell_threshold)); + _cell_clip.configure(&_cell_state_out1, nullptr, + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, + cell_threshold, -cell_threshold)); } // Configure block that calculates the output @@ -282,18 +384,20 @@ void NELSTMLayer::configure(const ITensor *input, _memory_group.manage(&_output1); _memory_group.manage(&_output4); - _fully_connected_output.configure(&_forget_gate_out2, &_output2, (_is_layer_norm_lstm) ? nullptr : output_gate_bias, &_output4); + _fully_connected_output.configure(&_forget_gate_out2, &_output2, (_is_layer_norm_lstm) ? nullptr : output_gate_bias, + &_output4); _output2.allocator()->allocate(); _forget_gate_out2.allocator()->allocate(); Tensor *output_gate_out = &_output4; - if(lstm_params.has_peephole_opt()) + 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(&_cell_state_out1, lstm_params.cell_to_output_weights(), &_output3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); + _pixelwise_mul_output_state1.configure(&_cell_state_out1, lstm_params.cell_to_output_weights(), &_output3, 1, + ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); _accum_output1.configure(&_output4, &_output3, &_output1, ConvertPolicy::SATURATE); _output4.allocator()->allocate(); output_gate_out = &_output1; @@ -305,21 +409,25 @@ void NELSTMLayer::configure(const ITensor *input, { _output1.allocator()->allocate(); } - if(_is_layer_norm_lstm) + 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(output_gate_out); - _pixelwise_mul_output_gate_coeff.configure(output_gate_out, lstm_params.output_layer_norm_weights(), &_output_layer_norm_out1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); + _pixelwise_mul_output_gate_coeff.configure(output_gate_out, lstm_params.output_layer_norm_weights(), + &_output_layer_norm_out1, 1, ConvertPolicy::SATURATE, + RoundingPolicy::TO_ZERO); // 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(&_output_layer_norm_out1, output_gate_bias, &_output_layer_norm_out2, ConvertPolicy::SATURATE); + _accum_output_gate_bias.configure(&_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(output_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); + _activation_output.configure(output_gate_out, nullptr, + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); // Configure block that calculates the output state /** lstm_res = PixelwiseMul(output, Activation(cell_state)) @@ -336,20 +444,24 @@ void NELSTMLayer::configure(const ITensor *input, _memory_group.manage(&_cell_state_activation); _activation_output_state.configure(&_cell_state_out1, &_cell_state_activation, activation_info); - _pixelwise_mul_output_state2.configure(&_cell_state_activation, output_gate_out, output_state_out_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); + _pixelwise_mul_output_state2.configure(&_cell_state_activation, output_gate_out, output_state_out_tmp, 1, + ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO); _cell_state_activation.allocator()->allocate(); output_gate_out->allocator()->allocate(); - if(lstm_params.has_projection()) + if (lstm_params.has_projection()) { _has_projection_weights = true; - _fully_connected_output_state.configure(output_state_out_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out); + _fully_connected_output_state.configure(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) + if (projection_threshold != 0.f) { _perform_projection_clipping = true; - _projection_clip.configure(output_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold)); + _projection_clip.configure(output_state_out, nullptr, + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, + -projection_threshold, projection_threshold)); } } @@ -359,7 +471,7 @@ void NELSTMLayer::configure(const ITensor *input, // Vector for holding the tensors to store in scratch buffer std::vector<const ITensor *> scratch_inputs; - if(!lstm_params.has_cifg_opt()) + if (!lstm_params.has_cifg_opt()) { scratch_inputs.emplace_back(input_gate_out); } @@ -373,29 +485,38 @@ void NELSTMLayer::configure(const ITensor *input, output_gate_out->allocator()->allocate(); } -Status NELSTMLayer::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<ITensorInfo> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold) +Status NELSTMLayer::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<ITensorInfo> &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); + 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); + 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); @@ -414,16 +535,16 @@ Status NELSTMLayer::validate(const ITensorInfo *input, 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)); + 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 (lstm_params.use_layer_norm()) { // If CIFG is used, input layer normalization weights tensor is omitted - if(lstm_params.has_cifg_opt()) + if (lstm_params.has_cifg_opt()) { ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights() != nullptr); } @@ -435,8 +556,12 @@ Status NELSTMLayer::validate(const ITensorInfo *input, 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_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); @@ -446,7 +571,7 @@ Status NELSTMLayer::validate(const ITensorInfo *input, } // Check peephole optimization - if(lstm_params.has_peephole_opt()) + 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); @@ -466,33 +591,39 @@ Status NELSTMLayer::validate(const ITensorInfo *input, std::vector<const ITensorInfo *> 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); + 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(NEConcatenateLayer::validate(inputs_vector, &forget_gate_concat, Window::DimX)); // Validate forget gate - ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, input_to_forget_weights, (lstm_params.use_layer_norm()) ? nullptr : forget_gate_bias, &forget_gate)); + ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate( + input, input_to_forget_weights, (lstm_params.use_layer_norm()) ? nullptr : forget_gate_bias, &forget_gate)); - if(lstm_params.has_peephole_opt()) + if (lstm_params.has_peephole_opt()) { - ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &forget_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); - ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &forget_gate, 1, + ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEArithmeticAddition::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE)); } - if(lstm_params.use_layer_norm()) + if (lstm_params.use_layer_norm()) { ARM_COMPUTE_RETURN_ON_ERROR(NEMeanStdDevNormalizationLayer::validate(&forget_gate)); - ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&forget_gate, lstm_params.forget_layer_norm_weights(), &forget_gate, 1, ConvertPolicy::SATURATE, - RoundingPolicy::TO_ZERO)); - ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&forget_gate, forget_gate_bias, &forget_gate, ConvertPolicy::SATURATE)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEPixelWiseMultiplication::validate(&forget_gate, lstm_params.forget_layer_norm_weights(), &forget_gate, 1, + ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEArithmeticAddition::validate(&forget_gate, forget_gate_bias, &forget_gate, ConvertPolicy::SATURATE)); } - ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&forget_gate, &forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); + ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate( + &forget_gate, &forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); // Validate input gate - if(!lstm_params.has_cifg_opt()) + 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()); + 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); @@ -500,88 +631,120 @@ Status NELSTMLayer::validate(const ITensorInfo *input, std::vector<const ITensorInfo *> 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()); + 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(NEConcatenateLayer::validate(lstm_weights, &lstm_gate_concat, Window::DimX)); - ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, lstm_params.input_to_input_weights(), (lstm_params.use_layer_norm()) ? nullptr : lstm_params.input_gate_bias(), &input_gate)); + ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::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()) + 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(NEPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_input_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); - ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&input_gate, &input_gate, &input_gate, ConvertPolicy::SATURATE)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_input_weights(), &input_gate, 1, + ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEArithmeticAddition::validate(&input_gate, &input_gate, &input_gate, ConvertPolicy::SATURATE)); } - if(lstm_params.use_layer_norm()) + if (lstm_params.use_layer_norm()) { ARM_COMPUTE_RETURN_ON_ERROR(NEMeanStdDevNormalizationLayer::validate(&input_gate)); - ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&input_gate, lstm_params.input_layer_norm_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); - ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&input_gate, lstm_params.input_gate_bias(), &input_gate, ConvertPolicy::SATURATE)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEPixelWiseMultiplication::validate(&input_gate, lstm_params.input_layer_norm_weights(), &input_gate, 1, + ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&input_gate, lstm_params.input_gate_bias(), + &input_gate, ConvertPolicy::SATURATE)); } - ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&input_gate, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); + ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate( + &input_gate, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); } else { - ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticSubtraction::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEArithmeticSubtraction::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE)); } // Validate cell state - ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, input_to_cell_weights, (lstm_params.use_layer_norm()) ? nullptr : cell_bias, &cell_state_tmp)); - ARM_COMPUTE_RETURN_ON_ERROR(NEGEMM::validate(output_state_in, &units_out_transposed_info, nullptr, &cell_state_tmp, 1.f, 0.f, GEMMInfo())); - ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE)); - if(lstm_params.use_layer_norm()) + ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate( + input, input_to_cell_weights, (lstm_params.use_layer_norm()) ? nullptr : cell_bias, &cell_state_tmp)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEGEMM::validate(output_state_in, &units_out_transposed_info, nullptr, &cell_state_tmp, 1.f, 0.f, GEMMInfo())); + ARM_COMPUTE_RETURN_ON_ERROR( + NEArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE)); + if (lstm_params.use_layer_norm()) { ARM_COMPUTE_RETURN_ON_ERROR(NEMeanStdDevNormalizationLayer::validate(&cell_state_tmp)); - ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&cell_state_tmp, lstm_params.cell_layer_norm_weights(), &cell_state_tmp, 1, ConvertPolicy::SATURATE, - RoundingPolicy::TO_ZERO)); - ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&cell_state_tmp, cell_bias, &cell_state_tmp, ConvertPolicy::SATURATE)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEPixelWiseMultiplication::validate(&cell_state_tmp, lstm_params.cell_layer_norm_weights(), &cell_state_tmp, + 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEArithmeticAddition::validate(&cell_state_tmp, cell_bias, &cell_state_tmp, ConvertPolicy::SATURATE)); } ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&cell_state_tmp, nullptr, activation_info)); - ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&cell_state_tmp, &input_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); - ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&cell_state_tmp, &forget_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); - ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE)); - if(cell_threshold != 0.f) + ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&cell_state_tmp, &input_gate, &cell_state_tmp, 1, + ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&cell_state_tmp, &forget_gate, &cell_state_tmp, 1, + ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE)); + if (cell_threshold != 0.f) { - ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&cell_state_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold, - cell_threshold))); + ARM_COMPUTE_RETURN_ON_ERROR( + NEActivationLayer::validate(&cell_state_tmp, nullptr, + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, + cell_threshold, -cell_threshold))); } // Validate output gate tmp std::vector<const ITensorInfo *> 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()); + 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(NEConcatenateLayer::validate(in_out_weights, &in_out_gate_concat, Window::DimX)); - ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, input_to_output_weights, (lstm_params.use_layer_norm()) ? nullptr : output_gate_bias, &output_gate_tmp)); + ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate( + input, input_to_output_weights, (lstm_params.use_layer_norm()) ? nullptr : output_gate_bias, &output_gate_tmp)); - if(lstm_params.has_peephole_opt()) + if (lstm_params.has_peephole_opt()) { - ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&cell_state_tmp, lstm_params.cell_to_output_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE, - RoundingPolicy::TO_ZERO)); - ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&output_gate_tmp, &output_gate_tmp, &output_gate_tmp, ConvertPolicy::SATURATE)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEPixelWiseMultiplication::validate(&cell_state_tmp, lstm_params.cell_to_output_weights(), &output_gate_tmp, + 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&output_gate_tmp, &output_gate_tmp, &output_gate_tmp, + ConvertPolicy::SATURATE)); } - if(lstm_params.use_layer_norm()) + if (lstm_params.use_layer_norm()) { ARM_COMPUTE_RETURN_ON_ERROR(NEMeanStdDevNormalizationLayer::validate(&output_gate_tmp)); - ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&output_gate_tmp, lstm_params.output_layer_norm_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE, - RoundingPolicy::TO_ZERO)); - ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&output_gate_tmp, output_gate_bias, &output_gate_tmp, ConvertPolicy::SATURATE)); + ARM_COMPUTE_RETURN_ON_ERROR( + NEPixelWiseMultiplication::validate(&output_gate_tmp, lstm_params.output_layer_norm_weights(), + &output_gate_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&output_gate_tmp, output_gate_bias, &output_gate_tmp, + ConvertPolicy::SATURATE)); } - ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&output_gate_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); + ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate( + &output_gate_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC))); // Validate output state ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&cell_state_tmp, &cell_state_tmp, activation_info)); - ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&cell_state_tmp, &output_gate_tmp, &output_gate_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); - if(lstm_params.has_projection()) + ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate( + &cell_state_tmp, &output_gate_tmp, &output_gate_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + if (lstm_params.has_projection()) { - ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::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(NEFullyConnectedLayer::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(NEActivationLayer::validate(output_state_out, output_state_out, - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold))); + ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate( + output_state_out, output_state_out, + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, + projection_threshold))); } } @@ -591,7 +754,7 @@ Status NELSTMLayer::validate(const ITensorInfo *input, // Validate scratch concatenation std::vector<const ITensorInfo *> inputs_vector_info_raw; - if(!lstm_params.has_cifg_opt()) + if (!lstm_params.has_cifg_opt()) { inputs_vector_info_raw.push_back(&input_gate); } @@ -612,12 +775,12 @@ void NELSTMLayer::run() _concat_inputs_forget_gate.run(); _fully_connected_forget_gate.run(); - if(_run_peephole_opt) + if (_run_peephole_opt) { _pixelwise_mul_forget_gate.run(); _accum_forget_gate1.run(); } - if(_is_layer_norm_lstm) + if (_is_layer_norm_lstm) { _mean_std_norm_forget_gate.run(); _pixelwise_mul_forget_gate_coeff.run(); @@ -625,15 +788,17 @@ void NELSTMLayer::run() } _activation_forget_gate.run(); - if(_run_cifg_opt) + if (_run_cifg_opt) { - if(_ones.info()->data_type() == DataType::F16) + if (_ones.info()->data_type() == DataType::F16) { - std::fill_n(reinterpret_cast<half *>(_ones.buffer()), _ones.info()->total_size() / _ones.info()->element_size(), 1); + std::fill_n(reinterpret_cast<half *>(_ones.buffer()), + _ones.info()->total_size() / _ones.info()->element_size(), 1); } else { - std::fill_n(reinterpret_cast<float *>(_ones.buffer()), _ones.info()->total_size() / _ones.info()->element_size(), 1); + std::fill_n(reinterpret_cast<float *>(_ones.buffer()), + _ones.info()->total_size() / _ones.info()->element_size(), 1); } _subtract_input_gate.run(); } @@ -641,13 +806,13 @@ void NELSTMLayer::run() { _fully_connected_input_gate.run(); - if(_run_peephole_opt) + if (_run_peephole_opt) { _pixelwise_mul_input_gate.run(); _accum_input_gate1.run(); } - if(_is_layer_norm_lstm) + if (_is_layer_norm_lstm) { _mean_std_norm_input_gate.run(); _pixelwise_mul_input_gate_coeff.run(); @@ -660,29 +825,30 @@ void NELSTMLayer::run() _transpose_cell_state.run(); _gemm_cell_state1.run(); _accum_cell_state1.run(); - if(_is_layer_norm_lstm) + 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) + if (_perform_cell_clipping) { _cell_clip.run(); } _fully_connected_output.run(); - if(_run_peephole_opt) + if (_run_peephole_opt) { _pixelwise_mul_output_state1.run(); _accum_output1.run(); } - if(_is_layer_norm_lstm) + if (_is_layer_norm_lstm) { _mean_std_norm_output_gate.run(); _pixelwise_mul_output_gate_coeff.run(); @@ -693,10 +859,10 @@ void NELSTMLayer::run() _activation_output_state.run(); _pixelwise_mul_output_state2.run(); - if(_has_projection_weights) + if (_has_projection_weights) { _fully_connected_output_state.run(); - if(_perform_projection_clipping) + if (_perform_projection_clipping) { _projection_clip.run(); } @@ -710,10 +876,10 @@ void NELSTMLayer::run() void NELSTMLayer::prepare() { - if(!_is_prepared) + if (!_is_prepared) { _concat_weights_forget_gate.run(); - if(!_run_cifg_opt) + if (!_run_cifg_opt) { _concat_weights_input_gate.run(); } |