Compute Library
 20.02.1
CLReduceMean.cpp
Go to the documentation of this file.
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  */
25 
29 #include "arm_compute/core/Error.h"
30 #include "arm_compute/core/Types.h"
33 
34 namespace arm_compute
35 {
36 namespace
37 {
38 Status validate_config(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
39 {
40  ARM_COMPUTE_UNUSED(keep_dims);
44  ARM_COMPUTE_RETURN_ERROR_ON(reduction_axis.num_dimensions() < 1);
45  ARM_COMPUTE_RETURN_ERROR_ON(reduction_axis.num_dimensions() > input->num_dimensions());
46 
47  const unsigned int reduction_ops = reduction_axis.num_dimensions();
48  const int input_dims = input->num_dimensions();
49  Coordinates axis_local = reduction_axis;
50 
51  for(unsigned int i = 0; i < axis_local.num_dimensions(); ++i)
52  {
53  //axis: The dimensions to reduce. Must be in the range [-rank(input_tensor), rank(input_tensor)).
54  ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] < (-static_cast<int>(input->num_dimensions())));
55  ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] >= static_cast<int>(input->num_dimensions()));
56  }
57 
58  if(output->tensor_shape().total_size() != 0)
59  {
60  // Only validate if not using auto_init for the output tensor
61  TensorShape out_shape = input->tensor_shape();
62  // Validate output_shape only if not using auto_init
63  convert_negative_axis(axis_local, input_dims);
64  std::sort(axis_local.begin(), axis_local.begin() + reduction_ops);
65  for(unsigned int i = 0; i < reduction_ops; ++i)
66  {
67  ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] > 3);
68  ARM_COMPUTE_RETURN_ERROR_ON(static_cast<unsigned int>(axis_local[i]) > input->num_dimensions() - 1);
69  if(output->total_size() > 0 && keep_dims)
70  {
71  ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(axis_local[i]) != 1);
72  }
73  if(keep_dims)
74  {
75  out_shape.set(axis_local[i], 1);
76  }
77  else
78  {
79  ARM_COMPUTE_RETURN_ERROR_ON(i > static_cast<unsigned int>(axis_local[i]));
80  const unsigned int remove_index = axis_local[i] - i;
81  ARM_COMPUTE_RETURN_ERROR_ON(remove_index >= out_shape.num_dimensions());
82  out_shape.remove_dimension(remove_index);
83  }
84  }
85  const TensorInfo out_info = input->clone()->set_tensor_shape(out_shape);
87  }
88  return Status{};
89 }
90 }
91 CLReduceMean::CLReduceMean(std::shared_ptr<IMemoryManager> memory_manager)
92  : _memory_group(std::move(memory_manager)), _reduction_kernels(), _reduced_outs(), _reshape(), _reduction_ops(), _keep_dims()
93 {
94 }
95 void CLReduceMean::configure(ICLTensor *input, const Coordinates &reduction_axis, bool keep_dims, ICLTensor *output)
96 {
97  // Perform validate step
98  ARM_COMPUTE_ERROR_THROW_ON(CLReduceMean::validate(input->info(), reduction_axis, keep_dims, output->info()));
99  // Output auto inizialitation if not yet initialized
101  auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
102 
103  _reduction_ops = reduction_axis.num_dimensions();
104  _reduction_kernels.resize(_reduction_ops);
105  _reduced_outs.resize(_reduction_ops - (keep_dims ? 1 : 0));
106  _keep_dims = keep_dims;
107 
108  Coordinates axis_local = reduction_axis;
109  const int input_dims = input->info()->num_dimensions();
110 
111  convert_negative_axis(axis_local, input_dims);
112 
113  // Perform reduction for every axis
114  for(int i = 0; i < _reduction_ops; ++i)
115  {
116  TensorShape out_shape = i == 0 ? input->info()->tensor_shape() : (&_reduced_outs[i - 1])->info()->tensor_shape();
117  out_shape.set(axis_local[i], 1);
118  auto in = (i == 0) ? input : (&_reduced_outs[i - 1]);
119 
120  if(i == _reduction_ops - 1 && keep_dims)
121  {
122  _reduction_kernels[i].configure(in, output, axis_local[i], ReductionOperation::MEAN_SUM);
123  }
124  else
125  {
126  _reduced_outs[i].allocator()->init(TensorInfo(out_shape, input->info()->num_channels(), input->info()->data_type(), input->info()->quantization_info()));
127  _memory_group.manage(&_reduced_outs[i]);
128  _reduction_kernels[i].configure(in, &_reduced_outs[i], axis_local[i], ReductionOperation::MEAN_SUM);
129  }
130  }
131 
132  // Allocate intermediate tensors
133  for(int i = 0; i < _reduction_ops - (keep_dims ? 1 : 0); ++i)
134  {
135  _reduced_outs[i].allocator()->allocate();
136  }
137 
138  // Configure reshape layer if we want to drop the dimensions
139  if(!keep_dims)
140  {
141  TensorShape out_shape = input->info()->tensor_shape();
142 
143  // We have to sort the reduction axis vectors in order for remove_dimension
144  // to work properly
145  std::sort(axis_local.begin(), axis_local.begin() + _reduction_ops);
146  for(int i = 0; i < _reduction_ops; ++i)
147  {
148  out_shape.remove_dimension(axis_local[i] - i);
149  }
150  auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(out_shape));
151  _reshape.configure(&_reduced_outs[_reduction_ops - 1], output);
152  }
153 }
154 
155 Status CLReduceMean::validate(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
156 {
157  return validate_config(input, reduction_axis, keep_dims, output);
158 }
159 
161 {
162  MemoryGroupResourceScope scope_mg(_memory_group);
163 
164  for(auto &kernel : _reduction_kernels)
165  {
166  kernel.run();
167  }
168 
169  if(!_keep_dims)
170  {
171  _reshape.run();
172  }
173 }
174 } // namespace arm_compute
#define ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(tensor)
Definition: CLValidate.h:34
Shape of a tensor.
Definition: TensorShape.h:39
void remove_dimension(size_t n)
Accessor to remove the dimension n from the tensor shape.
Definition: TensorShape.h:110
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:792
1 channel, 1 F32 per channel
Store the tensor's metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Status class.
Definition: Error.h:52
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
void configure(ICLTensor *input, const Coordinates &reduction_axis, bool keep_dims, ICLTensor *output)
Configure kernel.
Copyright (c) 2017-2020 ARM Limited.
bool auto_init_if_empty(ITensorInfo &info, const TensorShape &shape, int num_channels, DataType data_type, QuantizationInfo quantization_info=QuantizationInfo())
Auto initialize the tensor info (shape, number of channels and data type) if the current assignment i...
Definition: Helpers.inl:202
1 channel, 1 F16 per channel
void run() override
Run the kernels contained in the function.
void manage(IMemoryManageable *obj) override
Sets a object to be managed by the given memory group.
Definition: MemoryGroup.h:79
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
void run() override final
Run the kernels contained in the function.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(...)
Definition: Validate.h:443
quantized, asymmetric fixed-point 8-bit number unsigned
Coordinates of an item.
Definition: Coordinates.h:37
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
std::array< T, num_max_dimensions >::iterator begin()
Returns a read/write iterator that points to the first element in the dimension array.
Definition: Dimensions.h:194
static Status validate(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
Static function to check if given info will lead to a valid configuration of CLReduceMean.
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:163
Memory group resources scope handling class.
Definition: IMemoryGroup.h:82
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
unsigned int num_dimensions() const
Returns the effective dimensionality of the tensor.
Definition: Dimensions.h:122
Status validate_config(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
CLReduceMean(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Default constructor.
TensorShape & set(size_t dimension, size_t value, bool apply_dim_correction=true)
Accessor to set the value of one of the dimensions.
Definition: TensorShape.h:78
Store the tensor's metadata.
Definition: TensorInfo.h:45
void configure(const ICLTensor *input, ICLTensor *output)
Initialise the kernel's inputs and outputs.
quantized, asymmetric fixed-point 8-bit number signed
Coordinates & convert_negative_axis(Coordinates &coords, int max_value)
Convert negative coordinates to positive in the range [0, num_dims_input].
Definition: Helpers.h:774
TensorShape calculate_reduce_mean_shape(ITensor *input, const Coordinates &reduction_axis, bool keep_dims)
Calculate the output tensor shape for the reduce mean operation.