• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 "src/core/NEON/kernels/NEElementwiseUnaryKernel.h"
25 
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/ITensor.h"
29 #include "arm_compute/core/Validate.h"
30 #include "src/core/CPP/Validate.h"
31 #include "src/core/NEON/wrapper/wrapper.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33 #include "src/core/helpers/WindowHelpers.h"
34 #include "support/ToolchainSupport.h"
35 
36 namespace arm_compute
37 {
38 namespace
39 {
40 template <typename ScalarType>
elementwise_op_scalar_imp(ElementWiseUnary op,const ScalarType & a)41 inline ScalarType elementwise_op_scalar_imp(ElementWiseUnary op, const ScalarType &a)
42 {
43     switch(op)
44     {
45         case ElementWiseUnary::RSQRT:
46             return 1 / sqrt(a);
47         case ElementWiseUnary::EXP:
48             return std::exp(a);
49         case ElementWiseUnary::NEG:
50             return -a;
51         case ElementWiseUnary::LOG:
52             return std::log(a);
53         case ElementWiseUnary::ABS:
54             return std::abs(a);
55         case ElementWiseUnary::ROUND:
56             return support::cpp11::nearbyint(a);
57         case ElementWiseUnary::SIN:
58             return std::sin(a);
59         default:
60             ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
61     }
62 }
63 
64 template <typename ScalarType, typename VectorType>
elementwise_op_imp(ElementWiseUnary op,const VectorType & a)65 inline VectorType elementwise_op_imp(ElementWiseUnary op, const VectorType &a)
66 {
67     switch(op)
68     {
69         case ElementWiseUnary::RSQRT:
70             return wrapper::vinvsqrt(a);
71         case ElementWiseUnary::EXP:
72             return wrapper::vexpq(a);
73         case ElementWiseUnary::NEG:
74             return wrapper::vneg(a);
75         case ElementWiseUnary::LOG:
76             return wrapper::vlog(a);
77         case ElementWiseUnary::ABS:
78             return wrapper::vabs(a);
79         case ElementWiseUnary::ROUND:
80             return wrapper::vround(a);
81         case ElementWiseUnary::SIN:
82             return wrapper::vsin(a);
83         default:
84             ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
85     }
86 }
87 } // namespace
88 
89 template <typename ScalarType>
elementwise_op(const Window & window)90 void NEElementwiseUnaryKernel::elementwise_op(const Window &window)
91 {
92     const int  window_step_x  = 16 / sizeof(ScalarType);
93     const auto window_start_x = static_cast<int>(window.x().start());
94     const auto window_end_x   = static_cast<int>(window.x().end());
95 
96     Window win = window;
97     win.set(Window::DimX, Window::Dimension(0, 1, 1));
98 
99     Iterator input(_input, win);
100     Iterator output(_output, win);
101 
102     execute_window_loop(win, [&](const Coordinates &)
103     {
104         auto       output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
105         const auto input_ptr  = reinterpret_cast<const ScalarType *>(input.ptr());
106 
107         int x = window_start_x;
108         for(; x <= window_end_x - window_step_x; x += window_step_x)
109         {
110             wrapper::vstore(output_ptr + x, elementwise_op_imp<ScalarType>(_op, wrapper::vloadq(input_ptr + x)));
111         }
112         for(; x < window_end_x; ++x)
113         {
114             *(output_ptr + x) = elementwise_op_scalar_imp(_op, *(input_ptr + x));
115         }
116     },
117     input, output);
118 }
119 
NEElementwiseUnaryKernel()120 NEElementwiseUnaryKernel::NEElementwiseUnaryKernel()
121     : _func(nullptr), _input(nullptr), _output(nullptr), _op()
122 {
123 }
124 
configure(ElementWiseUnary op,const ITensor * input,ITensor * output)125 void NEElementwiseUnaryKernel::configure(ElementWiseUnary op, const ITensor *input, ITensor *output)
126 {
127     ARM_COMPUTE_ERROR_THROW_ON(validate(op, input->info(), output->info()));
128     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
129 
130     // Configure kernel window
131     const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input->info());
132     const TensorShape &out_shape    = broadcast_pair.first;
133     const ValidRegion &valid_region = broadcast_pair.second;
134 
135     // Auto initialize output if not initialized
136     auto_init_if_empty(*output->info(), out_shape, 1, input->info()->data_type());
137 
138     Window win = calculate_max_window(valid_region);
139 
140     _input  = input;
141     _output = output;
142     _op     = op;
143 
144     INEKernel::configure(win);
145 
146     switch(input->info()->data_type())
147     {
148         case DataType::F32:
149             _func = &NEElementwiseUnaryKernel::elementwise_op<float>;
150             break;
151 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
152         case DataType::F16:
153             _func = &NEElementwiseUnaryKernel::elementwise_op<float16_t>;
154 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
155             break;
156         case DataType::S32:
157             _func = &NEElementwiseUnaryKernel::elementwise_op<int32_t>;
158             break;
159         default:
160             ARM_COMPUTE_ERROR("DataType not supported");
161     }
162 }
163 
validate(ElementWiseUnary op,const ITensorInfo * input,const ITensorInfo * output)164 Status NEElementwiseUnaryKernel::validate(ElementWiseUnary op, const ITensorInfo *input, const ITensorInfo *output)
165 {
166     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
167     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
168     switch(op)
169     {
170         case ElementWiseUnary::EXP:
171         case ElementWiseUnary::RSQRT:
172         case ElementWiseUnary::LOG:
173         case ElementWiseUnary::ROUND:
174         case ElementWiseUnary::SIN:
175             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
176             break;
177         case ElementWiseUnary::NEG:
178         case ElementWiseUnary::ABS:
179             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32, DataType::S32);
180             break;
181         default:
182             ARM_COMPUTE_ERROR("ElementWiseUnary operation not supported");
183     }
184     // Validate in case of configured output
185     if(output->total_size() > 0)
186     {
187         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
188     }
189 
190     return Status{};
191 }
192 
run(const Window & window,const ThreadInfo & info)193 void NEElementwiseUnaryKernel::run(const Window &window, const ThreadInfo &info)
194 {
195     ARM_COMPUTE_UNUSED(info);
196     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
197     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
198     ARM_COMPUTE_ERROR_ON(_func == nullptr);
199     (this->*_func)(window);
200 }
201 } // namespace arm_compute
202