xref: /aosp_15_r20/external/ComputeLibrary/src/runtime/CL/functions/CLRNNLayer.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 #include "arm_compute/runtime/CL/functions/CLRNNLayer.h"
25 
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/Types.h"
28 #include "arm_compute/core/Utils.h"
29 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
30 #include "arm_compute/runtime/CL/CLScheduler.h"
31 #include "src/core/CL/kernels/CLFillBorderKernel.h"
32 
33 #include "src/common/utils/Log.h"
34 
35 namespace arm_compute
36 {
37 using namespace arm_compute::misc::shape_calculator;
38 
CLRNNLayer(std::shared_ptr<IMemoryManager> memory_manager)39 CLRNNLayer::CLRNNLayer(std::shared_ptr<IMemoryManager> memory_manager)
40     : _memory_group(std::move(memory_manager)), _gemm_state_f(), _add_kernel(), _activation(), _fully_connected_kernel(), _copy(), _fully_connected_out(), _gemm_output(), _add_output(),
41       _is_prepared(false)
42 {
43 }
44 
45 CLRNNLayer::~CLRNNLayer() = default;
46 
validate(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * recurrent_weights,const ITensorInfo * bias,const ITensorInfo * hidden_state,const ITensorInfo * output,const ActivationLayerInfo & info)47 Status CLRNNLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *recurrent_weights, const ITensorInfo *bias, const ITensorInfo *hidden_state,
48                             const ITensorInfo *output, const ActivationLayerInfo &info)
49 {
50     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, recurrent_weights, bias, hidden_state, output);
51     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(input, DataType::F16, DataType::F32);
52     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, recurrent_weights, bias, hidden_state, output);
53 
54     const int idx_width  = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
55     const int idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
56 
57     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != weights->dimension(idx_width));
58     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_height) != recurrent_weights->dimension(idx_width));
59     ARM_COMPUTE_RETURN_ERROR_ON(recurrent_weights->dimension(idx_width) != recurrent_weights->dimension(1));
60     ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() != 1);
61     ARM_COMPUTE_RETURN_ERROR_ON(bias->dimension(idx_width) != weights->dimension(idx_height));
62     ARM_COMPUTE_RETURN_ERROR_ON(hidden_state->dimension(idx_width) != weights->dimension(idx_height));
63     ARM_COMPUTE_RETURN_ERROR_ON(hidden_state->dimension(idx_height) != input->dimension(idx_height));
64     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), hidden_state->tensor_shape());
65 
66     auto shape_info = TensorInfo(compute_rnn_shape(recurrent_weights, hidden_state->dimension(idx_height)), 1, input->data_type());
67 
68     ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, weights, bias, &shape_info));
69     ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(hidden_state, recurrent_weights, nullptr, &shape_info, 1.f, 0.f));
70     ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&shape_info, &shape_info, &shape_info, ConvertPolicy::SATURATE));
71     ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&shape_info, &shape_info, info));
72 
73     return Status{};
74 }
75 
configure(const ICLTensor * input,const ICLTensor * weights,const ICLTensor * recurrent_weights,const ICLTensor * bias,ICLTensor * hidden_state,ICLTensor * output,ActivationLayerInfo & info)76 void CLRNNLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *recurrent_weights, const ICLTensor *bias, ICLTensor *hidden_state, ICLTensor *output,
77                            ActivationLayerInfo &info)
78 {
79     configure(CLKernelLibrary::get().get_compile_context(), input, weights, recurrent_weights, bias, hidden_state, output, info);
80 }
81 
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * weights,const ICLTensor * recurrent_weights,const ICLTensor * bias,ICLTensor * hidden_state,ICLTensor * output,ActivationLayerInfo & info)82 void CLRNNLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *recurrent_weights, const ICLTensor *bias,
83                            ICLTensor *hidden_state,
84                            ICLTensor *output, ActivationLayerInfo &info)
85 {
86     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, recurrent_weights, bias, hidden_state, output);
87     ARM_COMPUTE_ERROR_THROW_ON(CLRNNLayer::validate(input->info(), weights->info(), recurrent_weights->info(), bias->info(), hidden_state->info(), output->info(), info));
88     ARM_COMPUTE_LOG_PARAMS(input, weights, recurrent_weights, bias, hidden_state, output, info);
89 
90     const int   idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
91     TensorShape shape      = compute_rnn_shape(recurrent_weights->info(), hidden_state->info()->dimension(idx_height));
92 
93     _is_prepared = false;
94 
95     _fully_connected_out.allocator()->init(TensorInfo(shape, 1, input->info()->data_type()));
96     _gemm_output.allocator()->init(TensorInfo(shape, 1, input->info()->data_type()));
97 
98     // Manage intermediate buffers and configure
99     _memory_group.manage(&_fully_connected_out);
100     _fully_connected_kernel.configure(compile_context, input, weights, bias, &_fully_connected_out);
101 
102     _memory_group.manage(&_gemm_output);
103     _gemm_state_f.configure(compile_context, hidden_state, recurrent_weights, nullptr, &_gemm_output, 1.f, 0.f);
104 
105     _add_output.allocator()->init(TensorInfo(shape, 1, input->info()->data_type()));
106     _memory_group.manage(&_add_output);
107 
108     _add_kernel.configure(compile_context, &_fully_connected_out, &_gemm_output, &_add_output, ConvertPolicy::SATURATE);
109 
110     _fully_connected_out.allocator()->allocate();
111     _gemm_output.allocator()->allocate();
112 
113     _activation.configure(compile_context, &_add_output, hidden_state, info);
114     _add_output.allocator()->allocate();
115 
116     _copy.configure(compile_context, hidden_state, output);
117 }
118 
run()119 void CLRNNLayer::run()
120 {
121     prepare();
122 
123     MemoryGroupResourceScope scope_mg(_memory_group);
124 
125     _fully_connected_kernel.run();
126     _gemm_state_f.run();
127     _add_kernel.run();
128     _activation.run();
129 
130     // copy hidden out to output
131     _copy.run();
132 }
133 
prepare()134 void CLRNNLayer::prepare()
135 {
136     if(!_is_prepared)
137     {
138         _fully_connected_kernel.prepare();
139         _gemm_state_f.prepare();
140 
141         _is_prepared = true;
142     }
143 }
144 } // namespace arm_compute
145