xref: /aosp_15_r20/external/ComputeLibrary/src/runtime/NEON/functions/NERNNLayer.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2018-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 
25 #include "arm_compute/runtime/NEON/functions/NERNNLayer.h"
26 
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/core/TensorInfo.h"
29 #include "arm_compute/core/Types.h"
30 #include "arm_compute/core/Validate.h"
31 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
32 #include "arm_compute/runtime/NEON/NEScheduler.h"
33 #include "src/common/utils/Log.h"
34 
35 namespace arm_compute
36 {
37 NERNNLayer::~NERNNLayer() = default;
38 
NERNNLayer(std::shared_ptr<IMemoryManager> memory_manager)39 NERNNLayer::NERNNLayer(std::shared_ptr<IMemoryManager> memory_manager)
40     : _memory_group(std::move(memory_manager)), _gemm_state_f(), _add_f(), _activation(), _fully_connected(memory_manager), _copy_f(), _fully_connected_out(), _gemm_output(), _add_output(),
41       _is_prepared(false)
42 {
43 }
44 
validate(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * recurrent_weights,const ITensorInfo * bias,const ITensorInfo * hidden_state,const ITensorInfo * output,const ActivationLayerInfo & info)45 Status NERNNLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *recurrent_weights, const ITensorInfo *bias, const ITensorInfo *hidden_state,
46                             const ITensorInfo *output, const ActivationLayerInfo &info)
47 {
48     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, recurrent_weights, bias, hidden_state, output);
49     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(input, DataType::F16, DataType::F32);
50 
51     const int idx_width  = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
52     const int idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
53     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != weights->dimension(idx_width));
54     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() != 2);
55     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_height) != recurrent_weights->dimension(idx_width));
56     ARM_COMPUTE_RETURN_ERROR_ON(recurrent_weights->dimension(idx_width) != recurrent_weights->dimension(idx_height));
57     ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() != 1);
58     ARM_COMPUTE_RETURN_ERROR_ON(bias->dimension(idx_width) != weights->dimension(idx_height));
59     ARM_COMPUTE_RETURN_ERROR_ON(hidden_state->dimension(idx_width) != weights->dimension(idx_height));
60     ARM_COMPUTE_RETURN_ERROR_ON(hidden_state->dimension(idx_height) != input->dimension(idx_height));
61     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), hidden_state->tensor_shape());
62 
63     auto shape_info = TensorInfo(misc::shape_calculator::compute_rnn_shape(recurrent_weights, hidden_state->dimension(idx_height)), 1, input->data_type());
64 
65     ARM_COMPUTE_RETURN_ON_ERROR(NEFullyConnectedLayer::validate(input, weights, bias, &shape_info));
66     ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&shape_info, &shape_info, &shape_info, ConvertPolicy::SATURATE));
67     ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&shape_info, &shape_info, info));
68 
69     return Status{};
70 }
71 
configure(const ITensor * input,const ITensor * weights,const ITensor * recurrent_weights,const ITensor * bias,ITensor * hidden_state,ITensor * output,ActivationLayerInfo & info)72 void NERNNLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *recurrent_weights, const ITensor *bias, ITensor *hidden_state, ITensor *output,
73                            ActivationLayerInfo &info)
74 {
75     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, recurrent_weights, bias, hidden_state, output);
76     ARM_COMPUTE_ERROR_THROW_ON(NERNNLayer::validate(input->info(), weights->info(), recurrent_weights->info(), bias->info(), hidden_state->info(), output->info(), info));
77     ARM_COMPUTE_LOG_PARAMS(input, weights, recurrent_weights, bias, hidden_state, output, info);
78 
79     const int   idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
80     TensorShape shape      = misc::shape_calculator::compute_rnn_shape(recurrent_weights->info(), hidden_state->info()->dimension(idx_height));
81 
82     _is_prepared = false;
83 
84     // Manage intermediate buffers and configure
85     _fully_connected_out.allocator()->init(TensorInfo(shape, 1, input->info()->data_type()));
86     _gemm_output.allocator()->init(TensorInfo(shape, 1, input->info()->data_type()));
87 
88     // Manage intermediate buffers and configure
89     _memory_group.manage(&_fully_connected_out);
90     _fully_connected.configure(input, weights, bias, &_fully_connected_out);
91 
92     _memory_group.manage(&_gemm_output);
93     _gemm_state_f.configure(hidden_state, recurrent_weights, nullptr, &_gemm_output, 1.f, 0.f);
94 
95     _add_output.allocator()->init(TensorInfo(shape, 1, input->info()->data_type()));
96     _memory_group.manage(&_add_output);
97 
98     _add_f.configure(&_fully_connected_out, &_gemm_output, &_add_output, ConvertPolicy::SATURATE);
99 
100     _fully_connected_out.allocator()->allocate();
101     _gemm_output.allocator()->allocate();
102 
103     _activation.configure(&_add_output, hidden_state, info);
104     _add_output.allocator()->allocate();
105 
106     _copy_f.configure(hidden_state, output);
107 }
108 
run()109 void NERNNLayer::run()
110 {
111     prepare();
112 
113     MemoryGroupResourceScope scope_mg(_memory_group);
114 
115     _fully_connected.run();
116 
117     _gemm_state_f.run();
118 
119     _add_f.run();
120     _activation.run();
121 
122     // copy hidden out to output
123     _copy_f.run();
124 }
125 
prepare()126 void NERNNLayer::prepare()
127 {
128     if(!_is_prepared)
129     {
130         _fully_connected.prepare();
131         _gemm_state_f.prepare();
132 
133         _is_prepared = true;
134     }
135 }
136 } // namespace arm_compute
137