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/NESelectKernel.h"
25
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/IAccessWindow.h"
29 #include "arm_compute/core/ITensor.h"
30 #include "arm_compute/core/TensorInfo.h"
31 #include "arm_compute/core/Types.h"
32 #include "arm_compute/core/Validate.h"
33 #include "src/core/CPP/Validate.h"
34 #include "src/core/NEON/wrapper/wrapper.h"
35 #include "src/core/helpers/AutoConfiguration.h"
36 #include "src/core/helpers/WindowHelpers.h"
37 #include "utils/TypePrinter.h"
38
39 #include <arm_neon.h>
40 #include <map>
41 #include <string>
42
43 namespace arm_compute
44 {
45 namespace
46 {
47 template <typename ScalarType, typename VectorType>
select_op(const ITensor * cond,const ITensor * in1,const ITensor * in2,ITensor * out,const Window & window,const int window_step_x,const int window_start_x,const int window_end_x,const int limit,VectorType (* condition_conversion)(const uint8_t *))48 void select_op(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window,
49 const int window_step_x, const int window_start_x, const int window_end_x, const int limit, VectorType (*condition_conversion)(const uint8_t *))
50 {
51 Window win = window;
52 win.set(Window::DimX, Window::Dimension(0, 1, 1));
53
54 Iterator condition(cond, win);
55 Iterator input1(in1, win);
56 Iterator input2(in2, win);
57 Iterator output(out, win);
58
59 execute_window_loop(win, [&](const Coordinates &)
60 {
61 auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
62 const auto condition_ptr = reinterpret_cast<const uint8_t *>(condition.ptr());
63 const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr());
64 const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr());
65
66 int x = window_start_x;
67 for(; x <= limit; x += window_step_x)
68 {
69 const auto c = (*condition_conversion)(condition_ptr + x);
70 const auto a = wrapper::vloadq(input1_ptr + x);
71 const auto b = wrapper::vloadq(input2_ptr + x);
72 wrapper::vstore(output_ptr + x, wrapper::vbsl(c, a, b));
73 }
74 for(; x < window_end_x; ++x)
75 {
76 const auto c = *(condition_ptr + x);
77 const auto a = *(input1_ptr + x);
78 const auto b = *(input2_ptr + x);
79 *(output_ptr + x) = static_cast<bool>(c) ? a : b;
80 }
81 },
82 condition, input1, input2, output);
83 }
84
85 template <typename ScalarType, typename VectorType>
select_op_8(const ITensor * cond,const ITensor * in1,const ITensor * in2,ITensor * out,const Window & window)86 void select_op_8(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
87 {
88 const auto window_step_x = 16 / sizeof(ScalarType);
89 const auto window_start_x = static_cast<int>(window.x().start());
90 const auto window_end_x = static_cast<int>(window.x().end());
91
92 select_op<ScalarType, VectorType>(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr) -> VectorType
93 {
94 static const auto zero = wrapper::vdup_n(static_cast<uint8_t>(0), arm_compute::wrapper::traits::vector_128_tag());
95 return wrapper::vcgt(wrapper::vloadq(condition_ptr), zero);
96 });
97 }
98
99 template <typename ScalarType, typename VectorType>
select_op_16(const ITensor * cond,const ITensor * in1,const ITensor * in2,ITensor * out,const Window & window)100 void select_op_16(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
101 {
102 const auto window_step_x = 16 / sizeof(ScalarType);
103 const auto window_start_x = static_cast<int>(window.x().start());
104 const auto window_end_x = static_cast<int>(window.x().end());
105
106 select_op<ScalarType, VectorType>(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr) -> VectorType
107 {
108 static const auto zero = wrapper::vdup_n(static_cast<uint16_t>(0), arm_compute::wrapper::traits::vector_128_tag());
109 return wrapper::vcgt(wrapper::vmovl(wrapper::vload(condition_ptr)), zero);
110 });
111 }
112
113 template <typename ScalarType, typename VectorType>
select_op_32(const ITensor * cond,const ITensor * in1,const ITensor * in2,ITensor * out,const Window & window)114 void select_op_32(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
115 {
116 const auto window_step_x = 16 / sizeof(ScalarType);
117 const auto window_start_x = static_cast<int>(window.x().start());
118 const auto window_end_x = static_cast<int>(window.x().end());
119
120 select_op<ScalarType, VectorType>(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr) -> VectorType
121 {
122 static const auto zero = wrapper::vdup_n(static_cast<uint32_t>(0), arm_compute::wrapper::traits::vector_128_tag());
123 return wrapper::vcgt(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vload(condition_ptr)))), zero);
124 });
125 }
126
127 template <typename ScalarType>
select_op_not_same_rank(const ITensor * cond,const ITensor * in1,const ITensor * in2,ITensor * out,const Window & window)128 void select_op_not_same_rank(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window)
129 {
130 ARM_COMPUTE_UNUSED(window);
131
132 auto output_ptr = reinterpret_cast<ScalarType *>(out->buffer());
133 const auto condition_ptr = reinterpret_cast<const uint8_t *>(cond->buffer());
134 const auto input1_ptr = reinterpret_cast<const ScalarType *>(in1->buffer());
135 const auto input2_ptr = reinterpret_cast<const ScalarType *>(in2->buffer());
136
137 const int outer_size = cond->info()->total_size() / cond->info()->element_size();
138 const int inner_size = (in1->info()->total_size() / in1->info()->element_size()) / outer_size;
139 int offset = 0;
140 const int step = 16 / in1->info()->element_size();
141
142 for(int i = 0; i < outer_size; ++i)
143 {
144 int x = offset;
145 const auto input_ptr = static_cast<bool>(*(condition_ptr + i)) ? input1_ptr : input2_ptr;
146 for(; x <= offset + inner_size - step; x += step)
147 {
148 wrapper::vstore(output_ptr + x, wrapper::vloadq(input_ptr + x));
149 }
150 if(x <= offset + inner_size - (step / 2))
151 {
152 wrapper::vstore(output_ptr + x, wrapper::vload(input_ptr + x));
153 x += step / 2;
154 }
155 for(; x < offset + inner_size; ++x)
156 {
157 *(output_ptr + x) = *(input_ptr + x);
158 }
159 offset += inner_size;
160 }
161 }
162 } // namespace
163
NESelectKernel()164 NESelectKernel::NESelectKernel()
165 : _function(nullptr), _c(nullptr), _x(nullptr), _y(nullptr), _output(nullptr), _has_same_rank(false)
166 {
167 }
168
configure(const ITensor * c,const ITensor * x,const ITensor * y,ITensor * output)169 void NESelectKernel::configure(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output)
170 {
171 ARM_COMPUTE_ERROR_ON_NULLPTR(c, x, y, output);
172
173 // Auto initialize output if not initialized
174 auto_init_if_empty(*output->info(), x->info()->tensor_shape(), 1, x->info()->data_type());
175 ARM_COMPUTE_ERROR_THROW_ON(validate(c->info(), x->info(), y->info(), output->info()));
176
177 _c = c;
178 _x = x;
179 _y = y;
180 _output = output;
181 _has_same_rank = (c->info()->tensor_shape().num_dimensions() == x->info()->tensor_shape().num_dimensions());
182
183 std::string function_to_call("op_");
184 function_to_call += string_from_data_type(x->info()->data_type());
185
186 static std::map<std::string, SelectFunction *> map_function;
187
188 if(_has_same_rank)
189 {
190 map_function =
191 {
192 { "op_S8", &select_op_8<int8_t, uint8x16_t> },
193 { "op_S16", &select_op_16<int16_t, uint16x8_t> },
194 { "op_S32", &select_op_32<int32_t, uint32x4_t> },
195 { "op_U8", &select_op_8<uint8_t, uint8x16_t> },
196 { "op_U16", &select_op_16<uint16_t, uint16x8_t> },
197 { "op_U32", &select_op_32<uint32_t, uint32x4_t> },
198 { "op_F32", &select_op_32<float, uint32x4_t> }
199 };
200 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
201 map_function["op_F16"] = &select_op_16<float16_t, uint16x8_t>;
202 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
203 }
204 else
205 {
206 map_function =
207 {
208 { "op_S8", &select_op_not_same_rank<int8_t> },
209 { "op_S16", &select_op_not_same_rank<int16_t> },
210 { "op_S32", &select_op_not_same_rank<int32_t> },
211 { "op_U8", &select_op_not_same_rank<uint8_t> },
212 { "op_U16", &select_op_not_same_rank<uint16_t> },
213 { "op_U32", &select_op_not_same_rank<uint32_t> },
214 { "op_F32", &select_op_not_same_rank<float> }
215 };
216 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
217 map_function["op_F16"] = &select_op_not_same_rank<float16_t>;
218 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
219 }
220
221 auto it = map_function.find(function_to_call);
222
223 if(it != map_function.end())
224 {
225 _function = it->second;
226 }
227
228 Window win = calculate_max_window(x->info()->valid_region());
229 INEKernel::configure(win);
230 }
231
validate(const ITensorInfo * c,const ITensorInfo * x,const ITensorInfo * y,const ITensorInfo * output)232 Status NESelectKernel::validate(const ITensorInfo *c, const ITensorInfo *x, const ITensorInfo *y, const ITensorInfo *output)
233 {
234 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(c, x, y);
235 ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(x);
236 ARM_COMPUTE_RETURN_ERROR_ON(x->data_type() == DataType::UNKNOWN);
237 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(x, y);
238 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(x, y);
239 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(c, 1, DataType::U8);
240
241 const bool is_same_rank = (c->tensor_shape().num_dimensions() == x->tensor_shape().num_dimensions());
242 ARM_COMPUTE_RETURN_ERROR_ON(is_same_rank && (x->tensor_shape() != c->tensor_shape()));
243 ARM_COMPUTE_RETURN_ERROR_ON(!is_same_rank && ((c->tensor_shape().num_dimensions() > 1) || (c->tensor_shape().x() != x->tensor_shape()[x->tensor_shape().num_dimensions() - 1])));
244
245 if(output != nullptr && output->total_size() != 0)
246 {
247 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(x, output);
248 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(x, output);
249 }
250
251 return Status{};
252 }
253
run(const Window & window,const ThreadInfo & info)254 void NESelectKernel::run(const Window &window, const ThreadInfo &info)
255 {
256 ARM_COMPUTE_UNUSED(info);
257 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
258 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
259 ARM_COMPUTE_ERROR_ON(_function == nullptr);
260 _function(_c, _x, _y, _output, window);
261 }
262 } // namespace arm_compute
263