1 /*
2 * Copyright (c) 2018-2020 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/CLCopyKernel.h"
32 #include "src/core/CL/kernels/CLDepthConvertLayerKernel.h"
33 #include "src/core/CL/kernels/CLFillBorderKernel.h"
34 #include "src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.h"
35 #include "src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h"
36 #include "src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h"
37 #include "src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h"
38 #include "src/core/CL/kernels/CLGEMMLowpReductionKernel.h"
39 #include "src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h"
40 #include "src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h"
41 #include "src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h"
42 #include "src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
43 #include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
44 #include "support/MemorySupport.h"
45
46 namespace arm_compute
47 {
48 using namespace arm_compute::misc::shape_calculator;
49
CLRNNLayer(std::shared_ptr<IMemoryManager> memory_manager)50 CLRNNLayer::CLRNNLayer(std::shared_ptr<IMemoryManager> memory_manager)
51 : _memory_group(std::move(memory_manager)), _gemm_state_f(), _add_kernel(), _activation(), _fully_connected_kernel(), _copy_kernel(support::cpp14::make_unique<CLCopyKernel>()), _fully_connected_out(),
52 _gemm_output(), _add_output(), _is_prepared(false)
53 {
54 }
55
56 CLRNNLayer::~CLRNNLayer() = default;
57
validate(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * recurrent_weights,const ITensorInfo * bias,const ITensorInfo * hidden_state,const ITensorInfo * output,const ActivationLayerInfo & info)58 Status CLRNNLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *recurrent_weights, const ITensorInfo *bias, const ITensorInfo *hidden_state,
59 const ITensorInfo *output, const ActivationLayerInfo &info)
60 {
61 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, recurrent_weights, bias, hidden_state, output);
62 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(input, DataType::F16, DataType::F32);
63 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, recurrent_weights, bias, hidden_state, output);
64
65 const int idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
66 const int idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
67
68 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != weights->dimension(idx_width));
69 ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_height) != recurrent_weights->dimension(idx_width));
70 ARM_COMPUTE_RETURN_ERROR_ON(recurrent_weights->dimension(idx_width) != recurrent_weights->dimension(1));
71 ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() != 1);
72 ARM_COMPUTE_RETURN_ERROR_ON(bias->dimension(idx_width) != weights->dimension(idx_height));
73 ARM_COMPUTE_RETURN_ERROR_ON(hidden_state->dimension(idx_width) != weights->dimension(idx_height));
74 ARM_COMPUTE_RETURN_ERROR_ON(hidden_state->dimension(idx_height) != input->dimension(idx_height));
75 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), hidden_state->tensor_shape());
76
77 auto shape_info = TensorInfo(compute_rnn_shape(recurrent_weights, hidden_state->dimension(idx_height)), 1, input->data_type());
78
79 ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, weights, bias, &shape_info));
80 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(hidden_state, recurrent_weights, nullptr, &shape_info, 1.f, 0.f));
81 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&shape_info, &shape_info, &shape_info, ConvertPolicy::SATURATE));
82 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&shape_info, &shape_info, info));
83
84 return Status{};
85 }
86
configure(const ICLTensor * input,const ICLTensor * weights,const ICLTensor * recurrent_weights,const ICLTensor * bias,ICLTensor * hidden_state,ICLTensor * output,ActivationLayerInfo & info)87 void CLRNNLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *recurrent_weights, const ICLTensor *bias, ICLTensor *hidden_state, ICLTensor *output,
88 ActivationLayerInfo &info)
89 {
90 configure(CLKernelLibrary::get().get_compile_context(), input, weights, recurrent_weights, bias, hidden_state, output, info);
91 }
92
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)93 void CLRNNLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *recurrent_weights, const ICLTensor *bias,
94 ICLTensor *hidden_state,
95 ICLTensor *output, ActivationLayerInfo &info)
96 {
97 ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, recurrent_weights, bias, hidden_state, output);
98 ARM_COMPUTE_ERROR_THROW_ON(CLRNNLayer::validate(input->info(), weights->info(), recurrent_weights->info(), bias->info(), hidden_state->info(), output->info(), info));
99
100 const int idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
101 TensorShape shape = compute_rnn_shape(recurrent_weights->info(), hidden_state->info()->dimension(idx_height));
102
103 _is_prepared = false;
104
105 _fully_connected_out.allocator()->init(TensorInfo(shape, 1, input->info()->data_type()));
106 _gemm_output.allocator()->init(TensorInfo(shape, 1, input->info()->data_type()));
107
108 // Manage intermediate buffers and configure
109 _memory_group.manage(&_fully_connected_out);
110 _fully_connected_kernel.configure(compile_context, input, weights, bias, &_fully_connected_out);
111
112 _memory_group.manage(&_gemm_output);
113 _gemm_state_f.configure(compile_context, hidden_state, recurrent_weights, nullptr, &_gemm_output, 1.f, 0.f);
114
115 _add_output.allocator()->init(TensorInfo(shape, 1, input->info()->data_type()));
116 _memory_group.manage(&_add_output);
117
118 _add_kernel.configure(compile_context, &_fully_connected_out, &_gemm_output, &_add_output, ConvertPolicy::SATURATE);
119
120 _fully_connected_out.allocator()->allocate();
121 _gemm_output.allocator()->allocate();
122
123 _activation.configure(compile_context, &_add_output, hidden_state, info);
124 _add_output.allocator()->allocate();
125
126 _copy_kernel->configure(compile_context, hidden_state, output);
127 }
128
run()129 void CLRNNLayer::run()
130 {
131 prepare();
132
133 MemoryGroupResourceScope scope_mg(_memory_group);
134
135 _fully_connected_kernel.run();
136 _gemm_state_f.run();
137 _add_kernel.run();
138 _activation.run();
139
140 // copy hidden out to output
141 CLScheduler::get().enqueue(*_copy_kernel);
142 }
143
prepare()144 void CLRNNLayer::prepare()
145 {
146 if(!_is_prepared)
147 {
148 _fully_connected_kernel.prepare();
149 _gemm_state_f.prepare();
150
151 _is_prepared = true;
152 }
153 }
154 } // namespace arm_compute
155