xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2019-2022 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/NEON/kernels/NEMeanStdDevNormalizationKernel.h"
25 
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/ITensor.h"
28 #include "arm_compute/core/TensorInfo.h"
29 #include "arm_compute/core/Types.h"
30 #include "arm_compute/core/Window.h"
31 #include "src/core/CPP/Validate.h"
32 #include "src/core/NEON/NEMath.h"
33 #include "src/core/NEON/wrapper/wrapper.h"
34 #include "src/core/common/Registrars.h"
35 #include "src/core/helpers/AutoConfiguration.h"
36 #include "src/core/helpers/WindowHelpers.h"
37 #include "src/cpu/kernels/meanstddevnorm/list.h"
38 
39 namespace arm_compute
40 {
41 namespace
42 {
43 struct MeanStdDevNormSelectorData
44 {
45     DataType dt;
46 };
47 
48 using MeanStdDevNormSelctorPtr = std::add_pointer<bool(const MeanStdDevNormSelectorData &data)>::type;
49 using MeanStdDevNormUKernelPtr = std::add_pointer<void(ITensor *input, ITensor *output, float epsilon, const Window &window)>::type;
50 
51 struct MeanStdDevNormKernel
52 {
53     const char                    *name;
54     const MeanStdDevNormSelctorPtr is_selected;
55     MeanStdDevNormUKernelPtr       ukernel;
56 };
57 
58 static const std::vector<MeanStdDevNormKernel> available_kernels =
59 {
60     {
61         "fp32_neon_meanstddevnorm",
__anon9bfaead00202() 62         [](const MeanStdDevNormSelectorData & data) { return data.dt == DataType::F32; },
63         REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_meanstddevnorm)
64     },
65 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
66     {
67         "fp16_neon_meanstddevnorm",
__anon9bfaead00302() 68         [](const MeanStdDevNormSelectorData & data) { return data.dt == DataType::F16; },
69         REGISTER_FP16_NEON(arm_compute::cpu::neon_fp16_meanstddevnorm)
70     },
71 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
72     {
73         "qasymm8_neon_meanstddevnorm",
__anon9bfaead00402() 74         [](const MeanStdDevNormSelectorData & data) { return data.dt == DataType::QASYMM8; },
75         REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_meanstddevnorm)
76     },
77 };
78 
79 /** Micro-kernel selector
80  *
81  * @param[in] data Selection data passed to help pick the appropriate micro-kernel
82  *
83  * @return A matching micro-kernel else nullptr
84  */
get_implementation(const MeanStdDevNormSelectorData & data)85 const MeanStdDevNormKernel *get_implementation(const MeanStdDevNormSelectorData &data)
86 {
87     for(const auto &uk : available_kernels)
88     {
89         if(uk.is_selected(data))
90         {
91             return &uk;
92         }
93     }
94     return nullptr;
95 }
96 
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,float epsilon)97 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, float epsilon)
98 {
99     ARM_COMPUTE_UNUSED(epsilon);
100     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
101     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
102     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 2, "Input tensor cannot have more than 2 dimensions");
103     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32, DataType::QASYMM8);
104 
105     // Checks performed when output is configured
106     if((output != nullptr) && (output->total_size() != 0))
107     {
108         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
109         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
110     }
111     return Status{};
112 }
113 
validate_and_configure_window(ITensorInfo * input,ITensorInfo * output)114 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
115 {
116     if(output != nullptr)
117     {
118         ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
119         // Output auto inizialitation if not yet initialized
120         auto_init_if_empty(*output, *input);
121     }
122 
123     // This kernel doesn't need padding. A left-over for loop on dimension X, we cannot have any read or write out of memory
124     // For this reason num_elems_processed_per_iteration is set to 1
125     Window win = calculate_max_window(*input, Steps());
126 
127     return std::make_pair(Status{}, win);
128 }
129 } // namespace
130 
NEMeanStdDevNormalizationKernel()131 NEMeanStdDevNormalizationKernel::NEMeanStdDevNormalizationKernel()
132     : _input(nullptr), _output(nullptr), _epsilon(1e-8f)
133 {
134 }
135 
configure(ITensor * input,ITensor * output,float epsilon)136 void NEMeanStdDevNormalizationKernel::configure(ITensor *input, ITensor *output, float epsilon)
137 {
138     ARM_COMPUTE_ERROR_ON_NULLPTR(input);
139 
140     ARM_COMPUTE_ERROR_THROW_ON(NEMeanStdDevNormalizationKernel::validate(input->info(), (output != nullptr) ? output->info() : nullptr, epsilon));
141 
142     _input   = input;
143     _output  = (output == nullptr) ? input : output;
144     _epsilon = epsilon;
145 
146     // Configure kernel window
147     auto win_config = validate_and_configure_window(input->info(), (output == nullptr) ? nullptr : output->info());
148     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
149     ICPPKernel::configure(win_config.second);
150 }
151 
validate(const ITensorInfo * input,const ITensorInfo * output,float epsilon)152 Status NEMeanStdDevNormalizationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, float epsilon)
153 {
154     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, epsilon));
155     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (output != nullptr) ? output->clone().get() : nullptr).first);
156     return Status{};
157 }
158 
run(const Window & window,const ThreadInfo & info)159 void NEMeanStdDevNormalizationKernel::run(const Window &window, const ThreadInfo &info)
160 {
161     ARM_COMPUTE_UNUSED(info);
162     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
163     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
164 
165     const auto *uk = get_implementation(MeanStdDevNormSelectorData{ _output->info()->data_type() });
166     ARM_COMPUTE_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
167 
168     uk->ukernel(_input, _output, _epsilon, window);
169 }
170 } // namespace arm_compute
171