1 /*
2 * Copyright (c) 2017-2021 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/CLL2NormalizeLayerKernel.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 "arm_compute/core/Validate.h"
33 #include "src/core/CL/CLValidate.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "src/core/helpers/WindowHelpers.h"
36
37 #include "support/StringSupport.h"
38
39 namespace arm_compute
40 {
41 namespace
42 {
43 constexpr int max_input_tensor_dim = 3;
44
validate_arguments(const ITensorInfo * input,const ITensorInfo * sum,const ITensorInfo * output,int axis,float epsilon)45 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon)
46 {
47 ARM_COMPUTE_UNUSED(epsilon);
48
49 const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim);
50 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, sum, output);
51 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, sum);
52 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
53 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
54 ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis > 2, "Actual axis greater than 2 is not supported");
55 ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis >= TensorShape::num_max_dimensions, "Actual normalization axis greater than max number of dimensions");
56
57 // Reduce shape on axis
58 TensorShape sum_shape = input->tensor_shape();
59 sum_shape.set(actual_axis, 1);
60 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(sum->tensor_shape(), sum_shape);
61
62 if(output->total_size() != 0)
63 {
64 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
65 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
66 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
67 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(input->tensor_shape(), output->tensor_shape());
68 }
69
70 return Status{};
71 }
72 } // namespace
73
CLL2NormalizeLayerKernel()74 CLL2NormalizeLayerKernel::CLL2NormalizeLayerKernel()
75 : _input(nullptr), _sum(nullptr), _output(nullptr), _actual_axis(0), _epsilon(1e-12)
76 {
77 _type = CLKernelType::ELEMENTWISE;
78 }
79
configure(const ICLTensor * input,const ICLTensor * sum,ICLTensor * output,int axis,float epsilon)80 void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor *sum, ICLTensor *output, int axis, float epsilon)
81 {
82 configure(CLKernelLibrary::get().get_compile_context(), input, sum, output, axis, epsilon);
83 }
84
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * sum,ICLTensor * output,int axis,float epsilon)85 void CLL2NormalizeLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *sum, ICLTensor *output, int axis, float epsilon)
86 {
87 ARM_COMPUTE_ERROR_ON_NULLPTR(input, sum, output);
88 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), sum->info(), output->info(), axis, epsilon));
89 auto padding_info = get_padding_info({ input, sum, output });
90
91 _input = input;
92 _sum = sum;
93 _output = output;
94 _actual_axis = wrap_around(axis, max_input_tensor_dim);
95 _epsilon = epsilon;
96
97 const unsigned int vec_size_x = adjust_vec_size(max_cl_vector_width / input->info()->element_size(), input->info()->dimension(0));
98 const int vec_size_x_leftovers = input->info()->dimension(0) % vec_size_x;
99
100 // Set build options
101 CLBuildOptions build_opts;
102 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
103 build_opts.add_option("-DVEC_SIZE_X=" + support::cpp11::to_string(vec_size_x));
104 build_opts.add_option("-DVEC_SIZE_LEFTOVER_X=" + support::cpp11::to_string(vec_size_x_leftovers));
105
106 // Create kernel
107 std::string kernel_name;
108 unsigned int idx = 0;
109 switch(_actual_axis)
110 {
111 case 0:
112 kernel_name = "l2_normalize_x";
113 idx = num_arguments_per_2D_tensor() * 3;
114 break;
115 case 1:
116 kernel_name = "l2_normalize_y";
117 idx = num_arguments_per_2D_tensor() * 3;
118 break;
119 case 2:
120 kernel_name = "l2_normalize_z";
121 idx = num_arguments_per_3D_tensor() * 3;
122 break;
123 default:
124 ARM_COMPUTE_ERROR("Axis not supported");
125 }
126 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
127
128 // Set epsilon argument
129 if(input->info()->data_type() == DataType::F32)
130 {
131 _kernel.setArg<cl_float>(idx, _epsilon);
132 }
133 else
134 {
135 _kernel.setArg<cl_half>(idx, _epsilon);
136 }
137
138 // Configure kernel window
139 Window win = calculate_max_window(*input->info(), Steps(vec_size_x));
140
141 // Output tensor auto initialization if not yet initialized
142 auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type());
143
144 ICLKernel::configure_internal(win);
145 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
146 }
147
validate(const ITensorInfo * input,const ITensorInfo * sum,const ITensorInfo * output,int axis,float epsilon)148 Status CLL2NormalizeLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon)
149 {
150 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, sum, output, axis, epsilon));
151 return Status{};
152 }
153
run(const Window & window,cl::CommandQueue & queue)154 void CLL2NormalizeLayerKernel::run(const Window &window, cl::CommandQueue &queue)
155 {
156 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
157 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
158
159 Window window_sum(window);
160
161 switch(_actual_axis)
162 {
163 case 0:
164 {
165 window_sum.set(Window::DimX, Window::Dimension(0, 0, 0));
166 Window in_slice = window.first_slice_window_2D();
167 Window sum_slice = window_sum.first_slice_window_2D();
168 do
169 {
170 unsigned int idx = 0;
171 add_2D_tensor_argument(idx, _input, in_slice);
172 add_2D_tensor_argument(idx, _sum, sum_slice);
173 add_2D_tensor_argument(idx, _output, in_slice);
174 enqueue(queue, *this, in_slice, lws_hint());
175 }
176 while(window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(sum_slice));
177 }
178 break;
179 case 1:
180 {
181 window_sum.set(Window::DimY, Window::Dimension(0, 0, 0));
182 Window in_slice = window.first_slice_window_2D();
183 Window sum_slice = window_sum.first_slice_window_2D();
184 do
185 {
186 unsigned int idx = 0;
187 add_2D_tensor_argument(idx, _input, in_slice);
188 add_2D_tensor_argument(idx, _sum, sum_slice);
189 add_2D_tensor_argument(idx, _output, in_slice);
190 enqueue(queue, *this, in_slice, lws_hint());
191 }
192 while(window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(sum_slice));
193 }
194 break;
195 case 2:
196 {
197 window_sum.set(Window::DimZ, Window::Dimension(0, 0, 0));
198 Window in_slice = window.first_slice_window_3D();
199 Window sum_slice = window_sum.first_slice_window_3D();
200 do
201 {
202 unsigned int idx = 0;
203 add_3D_tensor_argument(idx, _input, in_slice);
204 add_3D_tensor_argument(idx, _sum, sum_slice);
205 add_3D_tensor_argument(idx, _output, in_slice);
206 enqueue(queue, *this, in_slice, lws_hint());
207 }
208 while(window.slide_window_slice_3D(in_slice) && window.slide_window_slice_3D(sum_slice));
209 }
210 break;
211 default:
212 ARM_COMPUTE_ERROR("Not supported");
213 }
214 }
215 } // namespace arm_compute
216