xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/kernels/CLFuseBatchNormalizationKernel.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 "src/core/CL/kernels/CLFuseBatchNormalizationKernel.h"
25 
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/CLKernelLibrary.h"
28 #include "arm_compute/core/CL/ICLTensor.h"
29 #include "arm_compute/core/Helpers.h"
30 #include "arm_compute/core/TensorInfo.h"
31 #include "arm_compute/core/Utils.h"
32 #include "src/core/CL/CLValidate.h"
33 #include "src/core/helpers/AutoConfiguration.h"
34 #include "src/core/helpers/WindowHelpers.h"
35 
36 #include "support/StringSupport.h"
37 
38 namespace arm_compute
39 {
40 namespace
41 {
validate_arguments(const ITensorInfo * input_weights,const ITensorInfo * bn_mean,const ITensorInfo * bn_var,const ITensorInfo * fused_weights,const ITensorInfo * fused_bias,const ITensorInfo * input_bias,const ITensorInfo * bn_beta,const ITensorInfo * bn_gamma,float epsilon,FuseBatchNormalizationType fbn_type)42 Status validate_arguments(const ITensorInfo *input_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
43                           const ITensorInfo *fused_weights, const ITensorInfo *fused_bias,
44                           const ITensorInfo *input_bias, const ITensorInfo *bn_beta, const ITensorInfo *bn_gamma,
45                           float epsilon, FuseBatchNormalizationType fbn_type)
46 {
47     ARM_COMPUTE_UNUSED(epsilon);
48     ARM_COMPUTE_ERROR_ON_NULLPTR(input_weights, bn_mean, bn_var);
49     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input_weights);
50     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_weights, 1, DataType::F16, DataType::F32);
51     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, bn_var);
52     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, bn_mean, bn_var);
53     ARM_COMPUTE_RETURN_ERROR_ON(input_bias == nullptr && fused_bias == nullptr);
54     ARM_COMPUTE_RETURN_ERROR_ON(bn_mean->num_dimensions() > 1);
55 
56     if(fbn_type == FuseBatchNormalizationType::CONVOLUTION)
57     {
58         ARM_COMPUTE_RETURN_ERROR_ON(input_weights->dimension(3) != bn_mean->dimension(0));
59     }
60     else
61     {
62         const size_t channel_idx = get_data_layout_dimension_index(input_weights->data_layout(), DataLayoutDimension::CHANNEL);
63         ARM_COMPUTE_RETURN_ERROR_ON(input_weights->dimension(channel_idx) != bn_mean->dimension(0));
64     }
65 
66     // Validate bias
67     if(input_bias != nullptr)
68     {
69         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, input_bias);
70         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, input_bias);
71     }
72     // Validate beta
73     if(bn_beta != nullptr)
74     {
75         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, bn_beta);
76         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, bn_beta);
77     }
78     // Validate gamma
79     if(bn_gamma != nullptr)
80     {
81         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, bn_gamma);
82         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, bn_gamma);
83     }
84     // Validate output weights
85     if(fused_weights != nullptr && fused_weights->total_size() != 0)
86     {
87         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input_weights, fused_weights);
88         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input_weights, fused_weights);
89         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, fused_weights);
90     }
91     // Validate output bias
92     if(fused_bias != nullptr && fused_bias->total_size() != 0)
93     {
94         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, fused_bias);
95         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, fused_bias);
96     }
97 
98     return Status{};
99 }
100 } // namespace
101 
CLFuseBatchNormalizationKernel()102 CLFuseBatchNormalizationKernel::CLFuseBatchNormalizationKernel()
103     : _input_weights(nullptr), _input_bias(nullptr), _bn_mean(nullptr), _bn_var(nullptr), _bn_gamma(nullptr), _bn_beta(nullptr), _fused_weights(nullptr), _fused_bias(nullptr), _epsilon(),
104       _run_in_place_weights(false), _run_in_place_bias(false)
105 {
106     _type = CLKernelType::ELEMENTWISE;
107 }
108 
configure(const ICLTensor * input_weights,const ICLTensor * bn_mean,const ICLTensor * bn_var,ICLTensor * fused_weights,ICLTensor * fused_bias,const ICLTensor * input_bias,const ICLTensor * bn_beta,const ICLTensor * bn_gamma,float epsilon,FuseBatchNormalizationType fbn_type)109 void CLFuseBatchNormalizationKernel::configure(const ICLTensor *input_weights, const ICLTensor *bn_mean, const ICLTensor *bn_var,
110                                                ICLTensor *fused_weights, ICLTensor *fused_bias,
111                                                const ICLTensor *input_bias, const ICLTensor *bn_beta, const ICLTensor *bn_gamma,
112                                                float epsilon, FuseBatchNormalizationType fbn_type)
113 {
114     configure(CLKernelLibrary::get().get_compile_context(), input_weights, bn_mean, bn_var, fused_weights, fused_bias, input_bias, bn_beta, bn_gamma, epsilon, fbn_type);
115 }
116 
configure(const CLCompileContext & compile_context,const ICLTensor * input_weights,const ICLTensor * bn_mean,const ICLTensor * bn_var,ICLTensor * fused_weights,ICLTensor * fused_bias,const ICLTensor * input_bias,const ICLTensor * bn_beta,const ICLTensor * bn_gamma,float epsilon,FuseBatchNormalizationType fbn_type)117 void CLFuseBatchNormalizationKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input_weights, const ICLTensor *bn_mean, const ICLTensor *bn_var,
118                                                ICLTensor *fused_weights, ICLTensor *fused_bias,
119                                                const ICLTensor *input_bias, const ICLTensor *bn_beta, const ICLTensor *bn_gamma,
120                                                float epsilon, FuseBatchNormalizationType fbn_type)
121 {
122     ARM_COMPUTE_ERROR_ON_NULLPTR(input_weights, bn_mean, bn_var);
123 
124     auto padding_info = get_padding_info({ input_weights, bn_mean, bn_var, fused_weights, fused_bias, input_bias, bn_beta, bn_gamma });
125 
126     _input_weights = input_weights;
127     _input_bias    = input_bias;
128     _bn_mean       = bn_mean;
129     _bn_var        = bn_var;
130     _bn_beta       = bn_beta;
131     _bn_gamma      = bn_gamma;
132     _fused_weights = fused_weights;
133     _fused_bias    = fused_bias;
134     _epsilon       = epsilon;
135 
136     _run_in_place_weights = (fused_weights == nullptr) || (fused_weights == input_weights);
137     _run_in_place_bias    = (input_bias != nullptr && fused_bias == nullptr) || (input_bias != nullptr && fused_bias == input_bias);
138 
139     // Auto initialize outputs
140     if(_fused_weights != nullptr)
141     {
142         // Output tensor auto initialization if not yet initialized
143         auto_init_if_empty(*_fused_weights->info(), *_input_weights->info()->clone());
144     }
145     if(_fused_bias != nullptr)
146     {
147         // Output tensor auto initialization if not yet initialized
148         auto_init_if_empty(*_fused_bias->info(), *_bn_mean->info()->clone());
149     }
150 
151     // Validate arguments
152     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input_weights->info(), bn_mean->info(), bn_var->info(),
153                                                   (fused_weights != nullptr) ? fused_weights->info() : nullptr,
154                                                   (fused_bias != nullptr) ? fused_bias->info() : nullptr,
155                                                   (input_bias != nullptr) ? input_bias->info() : nullptr,
156                                                   (bn_beta != nullptr) ? bn_beta->info() : nullptr,
157                                                   (bn_gamma != nullptr) ? bn_gamma->info() : nullptr,
158                                                   epsilon, fbn_type));
159 
160     // Configure kernel window
161     Window win = calculate_max_window(*input_weights->info());
162     ICLKernel::configure_internal(win);
163 
164     // Set build options
165     CLBuildOptions build_opts;
166     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input_weights->info()->data_type()));
167     build_opts.add_option_if(fbn_type == FuseBatchNormalizationType::CONVOLUTION, "-DDIM2=" + support::cpp11::to_string(input_weights->info()->dimension(2)));
168     build_opts.add_option("-DEPSILON=" + float_to_string_with_full_precision(epsilon));
169     build_opts.add_option_if(_input_weights->info()->data_layout() == DataLayout::NHWC, "-DNHWC");
170     build_opts.add_option_if(_run_in_place_weights, "-DIN_PLACE_W");
171     build_opts.add_option_if(_run_in_place_bias, "-DIN_PLACE_B");
172     build_opts.add_option_if(input_bias != nullptr, "-DBIAS");
173     build_opts.add_option_if(bn_beta != nullptr, "-DBETA");
174     build_opts.add_option_if(bn_gamma != nullptr, "-DGAMMA");
175 
176     // Create kernel
177     _kernel = create_kernel(compile_context, "fuse_batchnormalization_layer", build_opts.options());
178 
179     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
180 }
181 
validate(const ITensorInfo * input_weights,const ITensorInfo * bn_mean,const ITensorInfo * bn_var,const ITensorInfo * fused_weights,const ITensorInfo * fused_bias,const ITensorInfo * input_bias,const ITensorInfo * bn_beta,const ITensorInfo * bn_gamma,float epsilon,FuseBatchNormalizationType fbn_type)182 Status CLFuseBatchNormalizationKernel::validate(const ITensorInfo *input_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
183                                                 const ITensorInfo *fused_weights, const ITensorInfo *fused_bias,
184                                                 const ITensorInfo *input_bias, const ITensorInfo *bn_beta, const ITensorInfo *bn_gamma,
185                                                 float epsilon, FuseBatchNormalizationType fbn_type)
186 {
187     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input_weights, bn_mean, bn_var, fused_weights, fused_bias, input_bias, bn_beta, bn_gamma, epsilon, fbn_type));
188     return Status{};
189 }
190 
run(const arm_compute::Window & window,cl::CommandQueue & queue)191 void CLFuseBatchNormalizationKernel::run(const arm_compute::Window &window, cl::CommandQueue &queue)
192 {
193     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
194     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
195 
196     // Create window slice
197     Window collapsed_window = window.collapse(window, Window::DimZ);
198     Window slice_1d         = window.first_slice_window_1D();
199     Window slice_3d         = collapsed_window.first_slice_window_3D();
200 
201     // Add kernel arguments
202     unsigned int idx = 0;
203     add_3D_tensor_argument(idx, _input_weights, slice_3d);
204     if(_input_bias != nullptr)
205     {
206         add_1D_tensor_argument(idx, _input_bias, slice_1d);
207     }
208     add_1D_tensor_argument(idx, _bn_mean, slice_1d);
209     add_1D_tensor_argument(idx, _bn_var, slice_1d);
210     if(!_run_in_place_weights)
211     {
212         add_3D_tensor_argument(idx, _fused_weights, slice_3d);
213     }
214     if(!_run_in_place_bias)
215     {
216         add_1D_tensor_argument(idx, _fused_bias, slice_1d);
217     }
218     if(_bn_beta != nullptr)
219     {
220         add_1D_tensor_argument(idx, _bn_beta, slice_1d);
221     }
222     if(_bn_gamma != nullptr)
223     {
224         add_1D_tensor_argument(idx, _bn_gamma, slice_1d);
225     }
226     enqueue(queue, *this, slice_3d, lws_hint());
227 }
228 } // namespace arm_compute
229