1 /*
2 * Copyright (c) 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 "ClTemplateReshape.h"
25 #include "src/core/helpers/WindowHelpers.h"
26 #include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentGroup.h"
27
28 namespace arm_compute
29 {
30 namespace experimental
31 {
32 namespace dynamic_fusion
33 {
34 constexpr unsigned int vector_size_byte_opencl = 16;
35
ClTemplateReshape(ComponentId id,const ArgumentPack<ITensorInfo> & tensors)36 ClTemplateReshape::ClTemplateReshape(ComponentId id,
37 const ArgumentPack<ITensorInfo> &tensors)
38 : IGpuTemplateComponentWriter{ id, tensors },
39 _src{},
40 _dst{}
41 {
42 _src = this->tensors().get_const_tensor(TensorType::ACL_SRC_0);
43 _dst = this->tensors().get_const_tensor(TensorType::ACL_DST_0);
44 ARM_COMPUTE_ERROR_ON_NULLPTR(_src, _dst);
45 }
46
get_name() const47 std::string ClTemplateReshape::get_name() const
48 {
49 return "reshape";
50 }
51
get_component_code(const ComponentGroup & comp_group) const52 std::string ClTemplateReshape::get_component_code(const ComponentGroup &comp_group) const
53 {
54 ARM_COMPUTE_UNUSED(comp_group);
55 std::string code;
56
57 code = R"_(
58 //------------------ START KERNEL {{meta_kernel_id}} ---------------------
59
60 // IN(src) {{src}}
61 // OUT(dst, accum) {{dst}}
62
63 TILE(uint, M0, 1, g_dst_indirect_y);
64 {
65 __global uchar * base_src_ptr = {{src}}_ptr + {{src}}_offset_first_element_in_bytes;
66 const int tile_vertical_idx = g_ind_1 * {{arg_dst}}_c + g_ind_2 * {{arg_dst}}_c * {{arg_dst}}_w;
67 LOOP_UNROLLING(int, _m0, 0, 1, M0,
68 {
69 const int row_idx = _m0 * {{arg_dst}}_c + tile_vertical_idx;
70 const int tile_horizontal_idx = g_ind_0 + row_idx;
71 LOOP_UNROLLING(int, _n0, 0, 1, N0,
72 {
73 {{src}}_ptr = base_src_ptr;
74 const int linear_idx = tile_horizontal_idx + _n0;
75 const int in_id_x = linear_idx % {{src}}_c;
76 const int in_id_y = (linear_idx / {{src}}_c) % {{src}}_w;
77 const int in_id_z = linear_idx / ({{src}}_c * {{src}}_w);
78 {{src}}_ptr += in_id_x * sizeof({{DATA_TYPE}}) + in_id_y * {{src}}_stride_y + in_id_z * {{src}}_stride_z;
79 {{dst}}[_m0].s[_n0] = *((__global {{DATA_TYPE}} *){{src}}_ptr);
80 })
81 })
82
83 LOOP_UNROLLING(int, i, 0, 1, M0,
84 {
85 g_dst_indirect_y[i].v = (uint)min((int)(g_ind_1 + i), (int)({{arg_dst}}_w) - 1);
86 g_dst_indirect_y[i].v += (int)(g_ind_2 % {{arg_dst}}_h) * (int)({{arg_dst}}_w);
87 g_dst_indirect_y[i].v += (int)(g_ind_2 / {{arg_dst}}_h) * (int)({{arg_dst}}_w * {{arg_dst}}_h);
88 })
89 }
90 //------------------ END KERNEL {{meta_kernel_id}} ---------------------
91 )_";
92 return code;
93 }
94
declare_variables(GpuKernelVariableTable & vtable,const ComponentGroup & comp_group) const95 void ClTemplateReshape::declare_variables(GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const
96 {
97 vtable.declare_variable(
98 comp_group,
99 _src,
100 GpuKernelArgumentInfo(common_tensor_type), // GpuKernelArgumentInfo::Type::Image_3D
101 "src");
102
103 vtable.declare_variable(
104 comp_group,
105 _dst,
106 GpuKernelArgumentInfo(common_tensor_type),
107 "dst");
108 }
109
get_tag_lut(const GpuKernelVariableTable & vtable,const ComponentGroup & comp_group) const110 TagLUT ClTemplateReshape::get_tag_lut(const GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const
111 {
112 ARM_COMPUTE_UNUSED(comp_group);
113 TagLUT lut{};
114
115 // Arguments and global shared variables
116 lut["src"] = vtable.get_variable(_src);
117 lut["dst"] = vtable.get_variable(_dst);
118 lut["arg_dst"] = vtable.get_variable(comp_group.get_any_dst_tensor());
119 lut["meta_kernel_id"] = id();
120 lut["DATA_TYPE"] = get_cl_type_from_data_type(_dst->data_type());
121
122 return lut;
123 }
124
get_build_options(const ComponentGroup & comp_group) const125 CLBuildOptions ClTemplateReshape::get_build_options(const ComponentGroup &comp_group) const
126 {
127 CLBuildOptions build_opts{};
128 const auto root_window = comp_group.get_root_component()->template_writer()->get_window();
129 const unsigned int n0 = root_window.x().step();
130 const unsigned int m0 = root_window.y().step();
131 const unsigned int partial_store_n0 = _dst->dimension(0) % n0;
132 build_opts.add_option("-DN0=" + support::cpp11::to_string(n0));
133 build_opts.add_option("-DM0=" + support::cpp11::to_string(m0));
134 build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0));
135
136 return build_opts;
137 }
138
get_config_id() const139 std::string ClTemplateReshape::get_config_id() const
140 {
141 std::string config_id{};
142 config_id += lower_string(string_from_data_type(_dst->data_type()));
143 config_id += "_";
144 config_id += support::cpp11::to_string(_dst->dimension(0));
145 config_id += "_";
146 config_id += support::cpp11::to_string(_dst->dimension(1));
147
148 return config_id;
149 }
150
get_headers_list() const151 std::set<std::string> ClTemplateReshape::get_headers_list() const
152 {
153 return std::set<std::string>{ "helpers.h", "tile_helpers.h" };
154 }
155
get_window() const156 Window ClTemplateReshape::get_window() const
157 {
158 ARM_COMPUTE_ERROR_ON_MSG(_dst->tensor_shape().total_size() == 0U, "Destination tensor is not initialized");
159 const unsigned int n0 = adjust_vec_size(vector_size_byte_opencl / _dst->element_size(), _dst->dimension(0));
160 Window win = calculate_max_window(*_dst, Steps(n0));
161 return win.collapse(win, Window::DimZ);
162 }
163
164 } // namespace dynamic_fusion
165 } // namespace experimental
166 } // namespace arm_compute
167