1 /*
2  * Copyright (c) 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 
25 #include "ClTemplateResize.h"
26 
27 #include "src/core/helpers/WindowHelpers.h"
28 #include "src/core/utils/ScaleUtils.h"
29 #include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.h"
30 
31 namespace arm_compute
32 {
33 namespace experimental
34 {
35 namespace dynamic_fusion
36 {
ClTemplateResize(ComponentId id,const ArgumentPack<ITensorInfo> & tensors,const ClTemplateResize::Attributes & attributes)37 ClTemplateResize::ClTemplateResize(ComponentId id, const ArgumentPack<ITensorInfo> &tensors, const ClTemplateResize::Attributes &attributes)
38     : IGpuTemplateComponentWriter{ id, tensors }, _src{}, _dst{}, _attributes{ attributes }
39 {
40     _src = this->tensors().get_const_tensor(TensorType::ACL_SRC_0);
41     _dst = this->tensors().get_const_tensor(TensorType::ACL_DST_0);
42 
43     ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _dst);
44 }
45 
get_name() const46 std::string ClTemplateResize::get_name() const
47 {
48     return _attributes.interpolation_policy() == InterpolationPolicy::BILINEAR ? "resize_bilinear" : "resize_nearest";
49 }
50 
get_component_code(const IGpuTemplateComponentWriter::ComponentGroup & comp_group) const51 std::string ClTemplateResize::get_component_code(const IGpuTemplateComponentWriter::ComponentGroup &comp_group) const
52 {
53     ARM_COMPUTE_UNUSED(comp_group);
54 
55     std::string code = R"_(
56 //------------------ START KERNEL {{meta_kernel_id}} ---------------------
57 TILE(uint, 1, 1, g_dst_indirect_y);
58 {
59     const int yo = g_ind_2 % {{arg_dst}}_h;
60     const int bout = g_ind_2 / {{arg_dst}}_h;
61 )_";
62 
63     if(_attributes.interpolation_policy() == InterpolationPolicy::NEAREST_NEIGHBOR)
64     {
65         if(_attributes.sampling_policy() == SamplingPolicy::TOP_LEFT)
66         {
67             code += R"_(
68     float xi_f = (g_ind_1 * {{SCALE_X}});
69     float yi_f = (yo * {{SCALE_Y}});
70 )_";
71         }
72         else
73         {
74             code += R"_(
75     float xi_f = ((g_ind_1 + 0.5f) * {{SCALE_X}});
76     float yi_f = ((yo + 0.5f) * {{SCALE_Y}});
77 )_";
78         }
79 
80         if(_attributes.align_corners())
81         {
82             code += R"_(
83     xi_f = round(xi_f);
84     yi_f = round(yi_f);
85 )_";
86         }
87 
88         code += R"_(
89     const int xi0 = clamp((int)xi_f, 0, (int){{src}}_w - 1);
90     const int yi0 = clamp((int)yi_f, 0, (int){{src}}_h - 1);
91 
92     T_LOAD_NHWC_WITH_DILATION({{SRC_DATA_TYPE}}, 1, 1, N0, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yi0, xi0, g_ind_0, {{src}}_w, {{src}}_h, 1, 1, false, {{dst}});
93 )_";
94     }
95     else if(_attributes.interpolation_policy() == InterpolationPolicy::BILINEAR)
96     {
97         if(_attributes.sampling_policy() == SamplingPolicy::TOP_LEFT)
98         {
99             code += R"_(
100     float xi_f = (g_ind_1 * {{SCALE_X}});
101     float yi_f = (yo * {{SCALE_Y}});
102 )_";
103         }
104         else
105         {
106             code += R"_(
107     float xi_f = ((g_ind_1 + 0.5f) * {{SCALE_X}} - 0.5f);
108     float yi_f = ((yo + 0.5f) * {{SCALE_Y}} - 0.5f);
109 )_";
110         }
111 
112         code += R"_(
113     const int xi = (int)floor(xi_f);
114     const int yi = (int)floor(yi_f);
115 
116     TILE({{SRC_DATA_TYPE}}, 1, N0, in00);
117     TILE({{SRC_DATA_TYPE}}, 1, N0, in01);
118     TILE({{SRC_DATA_TYPE}}, 1, N0, in10);
119     TILE({{SRC_DATA_TYPE}}, 1, N0, in11);
120 
121     in00[0].v = {{CONSTANT_VALUE}};
122     in01[0].v = {{CONSTANT_VALUE}};
123     in10[0].v = {{CONSTANT_VALUE}};
124     in11[0].v = {{CONSTANT_VALUE}};
125 
126     const int xi0  = clamp(xi, 0, (int){{src}}_w - 1);
127     const int yi0  = clamp(yi, 0, (int){{src}}_h - 1);
128     const int xi1  = clamp(xi + 1, 0, (int){{src}}_w - 1);
129     const int yi1  = clamp(yi + 1, 0, (int){{src}}_h - 1);
130 
131     T_LOAD_NHWC_WITH_DILATION({{SRC_DATA_TYPE}}, 1, 1, N0, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yi0, xi0, g_ind_0, {{src}}_w, {{src}}_h, 1, 1, false, in00);
132     T_LOAD_NHWC_WITH_DILATION({{SRC_DATA_TYPE}}, 1, 1, N0, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yi0, xi1, g_ind_0, {{src}}_w, {{src}}_h, 1, 1, false, in01);
133     T_LOAD_NHWC_WITH_DILATION({{SRC_DATA_TYPE}}, 1, 1, N0, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yi1, xi0, g_ind_0, {{src}}_w, {{src}}_h, 1, 1, false, in10);
134     T_LOAD_NHWC_WITH_DILATION({{SRC_DATA_TYPE}}, 1, 1, N0, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yi1, xi1, g_ind_0, {{src}}_w, {{src}}_h, 1, 1, false, in11);
135 )_";
136 
137         if(is_data_type_float(_src->data_type()))
138         {
139             code += R"_(
140     const {{SRC_DATA_TYPE}} a  = ({{SRC_DATA_TYPE}})(xi_f - (float)xi);
141     const {{SRC_DATA_TYPE}} b  = ({{SRC_DATA_TYPE}})(1.f - a);
142     const {{SRC_DATA_TYPE}} a1 = ({{SRC_DATA_TYPE}})(yi_f - (float)yi);
143     const {{SRC_DATA_TYPE}} b1 = ({{SRC_DATA_TYPE}})(1.f - a1);
144 
145     // Calculate the output
146     {{dst}}[0].v = ((in00[0].v * b * b1) + (in01[0].v * a * b1) + (in10[0].v * b * a1) + (in11[0].v * a * a1));
147 )_";
148         }
149         else
150         {
151             code += R"_(
152     const float a  = (xi_f - (float)xi);
153     const float b  = (1.f - a);
154     const float a1 = (yi_f - (float)yi);
155     const float b1 = (1.f - a1);
156 
157     {{dst}}[0].v = CONVERT_SAT(
158         (CONVERT(in00[0].v, VEC_DATA_TYPE(float, N0)) * b * b1) +
159         (CONVERT(in01[0].v, VEC_DATA_TYPE(float, N0)) * a * b1) +
160         (CONVERT(in10[0].v, VEC_DATA_TYPE(float, N0)) * b * a1) +
161         (CONVERT(in11[0].v, VEC_DATA_TYPE(float, N0)) * a * a1), VEC_DATA_TYPE({{DST_DATA_TYPE}}, N0));
162 )_";
163         }
164     }
165     else
166     {
167         ARM_COMPUTE_ERROR("Unsupported interpolation policy");
168     }
169 
170     code += R"_(
171     g_dst_indirect_y[0].v = g_ind_1 + (yo * (int)({{arg_dst}}_w)) + bout * (int)({{arg_dst}}_w * {{arg_dst}}_h);
172 }
173 //------------------ END KERNEL {{meta_kernel_id}} ---------------------
174 )_";
175 
176     return code;
177 }
178 
declare_variables(GpuKernelVariableTable & vtable,const IGpuTemplateComponentWriter::ComponentGroup & comp_group) const179 void ClTemplateResize::declare_variables(GpuKernelVariableTable &vtable, const IGpuTemplateComponentWriter::ComponentGroup &comp_group) const
180 {
181     vtable.declare_variable(
182         comp_group,
183         _src,
184         GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer),
185         "src");
186 
187     vtable.declare_variable(
188         comp_group,
189         _dst,
190         GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer),
191         "dst");
192 }
193 
get_tag_lut(const GpuKernelVariableTable & vtable,const IGpuTemplateComponentWriter::ComponentGroup & comp_group) const194 TagLUT ClTemplateResize::get_tag_lut(const GpuKernelVariableTable &vtable, const IGpuTemplateComponentWriter::ComponentGroup &comp_group) const
195 {
196     TagLUT lut{};
197 
198     // Arguments and global shared variables
199     lut["src"] = vtable.get_variable(_src);
200     lut["dst"] = vtable.get_variable(_dst);
201 
202     const auto dst_argument = vtable.get_variable(comp_group.get_any_dst_tensor());
203     lut["arg_dst"]          = dst_argument.uniq_name;
204 
205     // Local build options
206     lut["meta_kernel_id"]  = id();
207     lut["SRC_DATA_TYPE"]   = get_cl_type_from_data_type(_src->data_type());
208     lut["SRC_TENSOR_TYPE"] = "BUFFER";
209     lut["DST_DATA_TYPE"]   = get_cl_type_from_data_type(_dst->data_type());
210     lut["CONSTANT_VALUE"]  = string_from_pixel_value(0, _src->data_type());
211 
212     const float scale_x = scale_utils::calculate_resize_ratio(_src->dimension(1), _dst->dimension(1), _attributes.align_corners());
213     const float scale_y = scale_utils::calculate_resize_ratio(_src->dimension(2), _dst->dimension(2), _attributes.align_corners());
214 
215     lut["SCALE_X"] = float_to_string_with_full_precision(scale_x);
216     lut["SCALE_Y"] = float_to_string_with_full_precision(scale_y);
217 
218     return lut;
219 }
220 
get_build_options(const IGpuTemplateComponentWriter::ComponentGroup & comp_group) const221 CLBuildOptions ClTemplateResize::get_build_options(const IGpuTemplateComponentWriter::ComponentGroup &comp_group) const
222 {
223     const Window       root_window = comp_group.get_root_component()->template_writer()->get_window();
224     const unsigned int n0          = root_window.x().step();
225     const unsigned int m0          = root_window.y().step();
226     const unsigned int partial_n0  = _dst->dimension(0) % n0;
227 
228     CLBuildOptions build_opts;
229 
230     build_opts.add_option("-DN0=" + support::cpp11::to_string(n0));
231     build_opts.add_option("-DM0=" + support::cpp11::to_string(m0));
232     build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_n0));
233 
234     return build_opts;
235 }
236 
get_config_id() const237 std::string ClTemplateResize::get_config_id() const
238 {
239     std::string config_id{};
240 
241     config_id += "resize_";
242     config_id += (_attributes.interpolation_policy() == InterpolationPolicy::NEAREST_NEIGHBOR ? "NEAREST_NEIGHBOR" : "");
243     config_id += (_attributes.interpolation_policy() == InterpolationPolicy::BILINEAR ? "BILINEAR" : "");
244     config_id += "_";
245     config_id += (_attributes.sampling_policy() == SamplingPolicy::CENTER ? "center" : "topleft");
246     config_id += "_";
247     config_id += support::cpp11::to_string(_dst->dimension(0));
248     config_id += "_";
249     config_id += support::cpp11::to_string(_dst->dimension(1));
250     config_id += "_";
251     config_id += support::cpp11::to_string(_dst->dimension(2));
252     config_id += "_";
253     config_id += support::cpp11::to_string(_dst->dimension(3));
254 
255     return config_id;
256 }
257 
get_headers_list() const258 std::set<std::string> ClTemplateResize::get_headers_list() const
259 {
260     return std::set<std::string>{ "helpers.h", "tile_helpers.h" };
261 }
262 
get_window() const263 Window ClTemplateResize::get_window() const
264 {
265     ARM_COMPUTE_ERROR_ON_MSG(_dst->tensor_shape().total_size() == 0U, "Destination tensor is not initialized");
266 
267     const unsigned int n0  = adjust_vec_size(16 / _src->element_size(), _src->dimension(0));
268     Window             win = calculate_max_window(*_dst, Steps(n0));
269     return win.collapse(win, Window::DimZ);
270 }
271 
272 } // namespace dynamic_fusion
273 } // namespace experimental
274 } // namespace arm_compute
275