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 
25 #include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateLogits1DMaxShiftExpSum.h"
26 
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 {
37 namespace
38 {
39     constexpr unsigned int serial_vector_size = 8;
40 } // namespace
ClTemplateLogits1DMaxShiftExpSum(ComponentId id,const ArgumentPack<ITensorInfo> & tensors,const Attributes & attributes)41 ClTemplateLogits1DMaxShiftExpSum::ClTemplateLogits1DMaxShiftExpSum(ComponentId                      id,
42                                                                    const ArgumentPack<ITensorInfo> &tensors,
43                                                                    const Attributes                &attributes)
44     : IGpuTemplateComponentWriter{ id, tensors },
45       _src{},
46       _sum{},
47       _dst{},
48       _attributes{ attributes }
49 {
50     _src = this->tensors().get_const_tensor(TensorType::ACL_SRC_0);
51     _sum = this->tensors().get_const_tensor(TensorType::ACL_DST_0);
52     _dst = this->tensors().get_const_tensor(TensorType::ACL_DST_1);
53     ARM_COMPUTE_ERROR_ON_NULLPTR(_src);
54     ARM_COMPUTE_ERROR_ON_NULLPTR(_sum);
55     ARM_COMPUTE_ERROR_ON_NULLPTR(_dst);
56 }
57 
get_name() const58 std::string ClTemplateLogits1DMaxShiftExpSum::get_name() const
59 {
60     return "logits_1d_max_shift_exp_sum";
61 }
62 
get_component_code(const ComponentGroup & comp_group) const63 std::string ClTemplateLogits1DMaxShiftExpSum::get_component_code(const ComponentGroup &comp_group) const
64 {
65     ARM_COMPUTE_UNUSED(comp_group);
66 
67     std::string code = R"_(
68 //------------------ START KERNEL {{meta_kernel_id}} ---------------------
69 #define VEC_TYPE VEC_DATA_TYPE({{DATA_TYPE}}, N0)
70 #define SELECT_TYPE SELECT_VEC_DATA_TYPE({{DATA_TYPE}}, N0)
71 {
72     __global uchar *src_addr = {{src}}_ptr + {{src}}_offset_first_element_in_bytes + g_ind_1 * {{src}}_stride_y + g_ind_2 * {{src}}_stride_z;
73     __global uchar *dst_addr = {{dst}}_ptr + {{dst}}_offset_first_element_in_bytes + g_ind_1 * {{dst}}_stride_y + g_ind_2 * {{dst}}_stride_z;
74     Image sum = CONVERT_TENSOR3D_TO_IMAGE_STRUCT({{sum}});
75     VEC_TYPE max_val_vec = (VEC_TYPE)({{MINVAL}});
76 )_";
77 
78     const bool beta_defined = (_attributes.beta() != 1.f);
79 
80     if(beta_defined)
81     {
82         code += R"_(
83     VEC_TYPE beta = (VEC_TYPE){{BETA}};
84 )_";
85     }
86 
87     constexpr unsigned int _serial_vector_size = 8;
88     const unsigned int     reduction_dim_size  = _src->dimension(0);
89     const unsigned int     vector_size         = adjust_vec_size(_serial_vector_size, reduction_dim_size);
90     const bool             non_multiple_of_n0  = ((reduction_dim_size % vector_size) != 0);
91 
92     if(non_multiple_of_n0)
93     {
94         code += R"_(
95     VEC_TYPE data    = VLOAD(N0)(0, (__global {{DATA_TYPE}} *)src_addr);
96     SELECT_TYPE widx = (SELECT_TYPE)PARTIAL_N0 > VEC_OFFS(SELECT_DATA_TYPE({{DATA_TYPE}}), N0);
97     max_val_vec      = max(max_val_vec, select((VEC_TYPE)({{MINVAL}}), data, widx));
98 )_";
99     }
100 
101     code += R"_(
102     for(uint i = PARTIAL_N0; i < {{SRC_WIDTH}}; i += N0)
103     {
104         VEC_TYPE data = VLOAD(N0)(0, (__global {{DATA_TYPE}} *)(src_addr + i * sizeof({{DATA_TYPE}})));
105         max_val_vec   = max(data, max_val_vec);
106     }
107 
108     {{DATA_TYPE}} max_val = MAX_REDUCE(max_val_vec, N0);
109     VEC_TYPE sum1D = 0;
110 )_";
111 
112     if(non_multiple_of_n0)
113     {
114         code += R"_(
115     data -= max_val;
116 )_";
117         if(beta_defined)
118         {
119             code += R"_(
120     data *= beta;
121 )_";
122         }
123 
124         if(_attributes.is_log_softmax())
125         {
126             code += R"_(
127     VSTORE_PARTIAL(N0, PARTIAL_N0)
128     (data, 0, (__global {{DATA_TYPE}} *)dst_addr);
129     data = exp(data);
130     data = select(0, data, widx);
131 )_";
132         }
133         else
134         {
135             code += R"_(
136     data = exp(data);
137     data = select(0, data, widx);
138     VSTORE_PARTIAL(N0, PARTIAL_N0)
139     (data, 0, (__global {{DATA_TYPE}} *)dst_addr);
140 )_";
141         }
142 
143         code += R"_(
144     sum1D += data;
145 )_";
146     }
147     code += R"_(
148     for(uint i = PARTIAL_N0; i < {{SRC_WIDTH}}; i += N0)
149     {
150         VEC_TYPE data = VLOAD(N0)(0, (__global {{DATA_TYPE}} *)(src_addr + i * sizeof({{DATA_TYPE}})));
151         data -= max_val;
152 )_";
153 
154     if(beta_defined)
155     {
156         code += R"_(
157     data *= beta;
158 )_";
159     }
160 
161     if(_attributes.is_log_softmax())
162     {
163         code += R"_(
164     VSTORE(N0)
165     (data, 0, (__global {{DATA_TYPE}} *)(dst_addr + i * sizeof({{DATA_TYPE}})));
166     data = exp(data);
167 )_";
168     }
169     else
170     {
171         code += R"_(
172     data = exp(data);
173     VSTORE(N0)
174     (data, 0, (__global {{DATA_TYPE}} *)(dst_addr + i * sizeof({{DATA_TYPE}})));
175 )_";
176     }
177 
178     code += R"_(
179     sum1D += data;
180     }
181 )_";
182 
183     code += R"_(
184     *((__global {{DATA_TYPE}} *)sum.ptr) = SUM_REDUCE(sum1D, N0);
185 }
186 //------------------ END KERNEL {{meta_kernel_id}} ---------------------
187 )_";
188 
189     return code;
190 }
191 
declare_variables(GpuKernelVariableTable & vtable,const ComponentGroup & comp_group) const192 void ClTemplateLogits1DMaxShiftExpSum::declare_variables(GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const
193 {
194     vtable.declare_variable(
195         comp_group,
196         _src,
197         GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Tensor_3D),
198         "src");
199 
200     vtable.declare_variable(
201         comp_group,
202         _sum,
203         GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Tensor_3D),
204         "sum");
205 
206     vtable.declare_variable(
207         comp_group,
208         _dst,
209         GpuKernelArgumentInfo(GpuKernelArgumentInfo::Type::Tensor_3D),
210         "dst");
211 }
212 
get_tag_lut(const GpuKernelVariableTable & vtable,const ComponentGroup & comp_group) const213 TagLUT ClTemplateLogits1DMaxShiftExpSum::get_tag_lut(const GpuKernelVariableTable &vtable, const ComponentGroup &comp_group) const
214 {
215     ARM_COMPUTE_UNUSED(comp_group);
216 
217     TagLUT lut{};
218 
219     // Arguments and global shared variables
220     lut["src"] = vtable.get_variable(_src);
221     lut["sum"] = vtable.get_variable(_sum);
222     lut["dst"] = vtable.get_variable(_dst);
223 
224     // Local build options
225     lut["meta_kernel_id"] = id();
226 
227     const DataType data_type = _src->data_type();
228 
229     lut["DATA_TYPE"] = get_cl_type_from_data_type(data_type);
230     lut["BETA"]      = float_to_string_with_full_precision(_attributes.beta());
231     lut["MINVAL"]    = (data_type == DataType::F16) ? std::string("-HALF_MAX") : std::string("-FLT_MAX");
232     lut["SRC_WIDTH"] = support::cpp11::to_string(_src->dimension(0));
233 
234     return lut;
235 }
236 
get_build_options(const ComponentGroup & comp_group) const237 CLBuildOptions ClTemplateLogits1DMaxShiftExpSum::get_build_options(const ComponentGroup &comp_group) const
238 {
239     ARM_COMPUTE_UNUSED(comp_group);
240     CLBuildOptions build_opts{};
241 
242     const unsigned int     reduction_dim_size = _src->dimension(0);
243     const unsigned int     vector_size        = adjust_vec_size(serial_vector_size, reduction_dim_size);
244 
245     build_opts.add_option("-DN0=" + support::cpp11::to_string(vector_size));
246     build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string((reduction_dim_size % vector_size)));
247 
248     return build_opts;
249 }
250 
get_config_id() const251 std::string ClTemplateLogits1DMaxShiftExpSum::get_config_id() const
252 {
253     std::string config_id = get_name();
254 
255     config_id += "_";
256     config_id += support::cpp11::to_string(_src->dimension(0));
257     config_id += "_";
258     config_id += string_from_data_type(_src->data_type());
259 
260     return config_id;
261 }
262 
get_headers_list() const263 std::set<std::string> ClTemplateLogits1DMaxShiftExpSum::get_headers_list() const
264 {
265     return std::set<std::string>{ "helpers.h", "tile_helpers.h" };
266 }
267 
get_window() const268 Window ClTemplateLogits1DMaxShiftExpSum::get_window() const
269 {
270     ARM_COMPUTE_ERROR_ON_MSG(_dst->tensor_shape().total_size() == 0U, "Destination tensor is not initialized");
271 
272     Window win = calculate_max_window(*_dst, Steps(_src->dimension(0)));
273     return win.collapse(win, Window::DimZ);
274 }
275 
276 } // namespace dynamic_fusion
277 } // namespace experimental
278 } // namespace arm_compute
279