1 /*
2 * Copyright (c) 2022-2023 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 "ClTemplateActivation.h"
25
26 #include "arm_compute/core/Utils.h"
27 #include "src/core/helpers/WindowHelpers.h"
28 #include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.h"
29 #include "support/StringSupport.h"
30
31 namespace arm_compute
32 {
33 namespace experimental
34 {
35 namespace dynamic_fusion
36 {
ClTemplateActivation(ComponentId id,const ArgumentPack<ITensorInfo> & tensors,const Attributes & attributes)37 ClTemplateActivation::ClTemplateActivation(ComponentId id,
38 const ArgumentPack<ITensorInfo> &tensors,
39 const Attributes &attributes)
40 : IGpuTemplateComponentWriter{ id, tensors },
41 _src{},
42 _dst{},
43 _attributes{ attributes }
44 {
45 _src = this->tensors().get_const_tensor(TensorType::ACL_SRC);
46 _dst = this->tensors().get_const_tensor(TensorType::ACL_DST);
47 ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _dst);
48 }
49
get_name() const50 std::string ClTemplateActivation::get_name() const
51 {
52 return "activation";
53 }
54
get_component_code(const ComponentGroup & comp_group) const55 std::string ClTemplateActivation::get_component_code(const ComponentGroup &comp_group) const
56 {
57 std::string code;
58 const bool is_root = (comp_group.get_root_component()->id() == this->id());
59
60 code = R"_(
61 //------------------ START KERNEL {{meta_kernel_id}} ---------------------
62 )_";
63 if(is_root)
64 {
65 code += R"_(
66 // IN(src) {{src}}
67 // OUT(dst, accum) {{dst}}
68
69 TILE({{DATA_TYPE}}, M0, N0, {{src}});
70 TILE(uint, M0, 1, g_dst_indirect_y);
71 {
72 {{src}}_offset_first_element_in_bytes += g_ind_2 * {{src}}_stride_z;
73
74 T_LOAD({{DATA_TYPE}}, M0, N0, {{TENSOR_TYPE}}, {{src}}, g_ind_0, g_ind_1, 1, {{src}}_stride_y, {{src}});
75
76 T_ACTIVATION({{DATA_TYPE}}, M0, N0, {{ACT}}, {{A_VAL}}, {{B_VAL}}, {{src}}, {{dst}});
77 }
78
79 LOOP_UNROLLING(int, i, 0, 1, M0,
80 {
81 g_dst_indirect_y[i].v = (uint)min((int)(g_ind_1 + i), (int)({{arg_dst}}_w) - 1);
82 g_dst_indirect_y[i].v += (int)(g_ind_2 % {{arg_dst}}_h) * (int)({{arg_dst}}_w);
83 g_dst_indirect_y[i].v += (int)(g_ind_2 / {{arg_dst}}_h) * (int)({{arg_dst}}_w * {{arg_dst}}_h);
84 })
85 )_";
86 }
87 else
88 {
89 code += R"_(
90 // IN/OUT(src, accum) {{src}}
91
92 {
93 T_ACTIVATION({{DATA_TYPE}}, M0, N0, {{ACT}}, {{A_VAL}}, {{B_VAL}}, {{src}}, {{dst}});
94 }
95 )_";
96 }
97 code += R"_(
98 //------------------ END KERNEL {{meta_kernel_id}} ---------------------
99 )_";
100 return code;
101 }
102
declare_variables(GpuKernelVariableTable & vtable,const ComponentGroup & comp_group) const103 void ClTemplateActivation::declare_variables(GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const
104 {
105 vtable.declare_variable(
106 comp_group,
107 _src,
108 GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer),
109 "src");
110
111 vtable.declare_variable(
112 comp_group,
113 _dst,
114 GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Tensor_4D_t_Buffer),
115 "dst");
116 }
117
get_tag_lut(const GpuKernelVariableTable & vtable,const ComponentGroup & comp_group) const118 TagLUT ClTemplateActivation::get_tag_lut(const GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const
119 {
120 ARM_COMPUTE_UNUSED(comp_group);
121
122 TagLUT lut{};
123 // Arguments and global shared variables
124 lut["src"] = vtable.get_variable(_src);
125 lut["dst"] = vtable.get_variable(_dst);
126
127 const auto dst_argument = vtable.get_variable(comp_group.get_any_dst_tensor());
128 lut["arg_dst"] = dst_argument.uniq_name;
129
130 // Local build options
131 lut["meta_kernel_id"] = id();
132 lut["DATA_TYPE"] = get_cl_type_from_data_type(_src->data_type());
133 lut["TENSOR_TYPE"] = "BUFFER";
134
135 const auto f_act = lower_string(string_from_activation_func(_attributes.activation()));
136
137 lut["ACT"] = f_act;
138 lut["A_VAL"] = float_to_string_with_full_precision(_attributes.a());
139 lut["B_VAL"] = float_to_string_with_full_precision(_attributes.b());
140
141 return lut;
142 }
143
get_build_options(const ComponentGroup & comp_group) const144 CLBuildOptions ClTemplateActivation::get_build_options(const ComponentGroup &comp_group) const
145 {
146 /// NOTE: For now tile sizes (n0, m0) are set by the execution window. This may change in the future
147 const auto root_window = comp_group.get_root_component()->template_writer()->get_window();
148 const unsigned int n0 = root_window.x().step();
149 const unsigned int m0 = root_window.y().step();
150 const unsigned int partial_store_n0 = _dst->dimension(0) % n0;
151
152 CLBuildOptions build_opts;
153 build_opts.add_option("-DN0=" + support::cpp11::to_string(n0));
154 build_opts.add_option("-DM0=" + support::cpp11::to_string(m0));
155 build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0));
156
157 return build_opts;
158 }
159
get_config_id() const160 std::string ClTemplateActivation::get_config_id() const
161 {
162 std::string config_id{};
163 config_id += "activation_";
164 config_id += lower_string(string_from_data_type(_src->data_type()));
165 config_id += "_";
166 config_id += support::cpp11::to_string(_src->dimension(0));
167 config_id += "_";
168 config_id += support::cpp11::to_string(_src->dimension(1));
169 return config_id;
170 }
171
get_headers_list() const172 std::set<std::string> ClTemplateActivation::get_headers_list() const
173 {
174 return std::set<std::string>{ "helpers.h", "tile_helpers.h", "activation_float_helpers.h" };
175 }
176
get_window() const177 Window ClTemplateActivation::get_window() const
178 {
179 ARM_COMPUTE_ERROR_ON_MSG(_dst->tensor_shape().total_size() == 0U, "Destination tensor is not initialized");
180 const unsigned int n0 = adjust_vec_size(16 / _dst->element_size(), _dst->dimension(0));
181 Window win = calculate_max_window(*_dst, Steps(n0));
182 return win.collapse(win, Window::DimZ);
183 }
184
185 } // namespace dynamic_fusion
186 } // namespace experimental
187 } // namespace arm_compute
188