xref: /aosp_15_r20/external/ComputeLibrary/src/gpu/cl/kernels/ClActivationKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2016-2021 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/gpu/cl/kernels/ClActivationKernel.h"
25 
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/ICLTensor.h"
28 #include "arm_compute/core/TensorInfo.h"
29 #include "arm_compute/core/Utils.h"
30 #include "src/core/CL/CLValidate.h"
31 #include "src/core/helpers/AutoConfiguration.h"
32 #include "src/core/helpers/WindowHelpers.h"
33 #include "support/Cast.h"
34 
35 #include "support/StringSupport.h"
36 
37 #include <set>
38 
39 namespace arm_compute
40 {
41 namespace opencl
42 {
43 namespace kernels
44 {
45 namespace
46 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * dst,const ActivationLayerInfo & act_info)47 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const ActivationLayerInfo &act_info)
48 {
49     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
50     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16, DataType::F16, DataType::F32);
51 
52     static std::set<ActivationLayerInfo::ActivationFunction> quantized_supported_activations =
53     {
54         ActivationLayerInfo::ActivationFunction::RELU,
55         ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU,
56         ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
57         ActivationLayerInfo::ActivationFunction::LOGISTIC,
58         ActivationLayerInfo::ActivationFunction::TANH,
59         ActivationLayerInfo::ActivationFunction::HARD_SWISH,
60         ActivationLayerInfo::ActivationFunction::LEAKY_RELU,
61     };
62     const DataType                                data_type = src->data_type();
63     const QuantizationInfo                       &oq_info   = (dst != nullptr) ? dst->quantization_info() : src->quantization_info();
64     const ActivationLayerInfo::ActivationFunction f_act     = act_info.activation();
65 
66     ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_quantized(data_type) && (quantized_supported_activations.count(f_act) == 0),
67                                     "For Quantized data type only hard swish, leaky relu, tanh, logistic, relu and lower/upper bounded relu are supported");
68 
69     ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8 && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 128.f, 128)));
70     ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8 && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 256.f, 0)));
71 
72     ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_symmetric(data_type) && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 32768.f, 0)));
73     ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_symmetric(data_type) && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 32768.f, 0)));
74 
75     ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8_SIGNED && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 128.f, 0)));
76     ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8_SIGNED && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 256.f, -128)));
77 
78     // Checks performed when destination is configured
79     if((dst != nullptr) && (dst->total_size() != 0))
80     {
81         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
82         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
83     }
84 
85     return Status{};
86 }
87 } // namespace
88 
ClActivationKernel()89 ClActivationKernel::ClActivationKernel()
90 {
91     _type = CLKernelType::ELEMENTWISE;
92 }
93 
configure(const ClCompileContext & compile_context,ITensorInfo * src,ITensorInfo * dst,ActivationLayerInfo act_info)94 void ClActivationKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, ActivationLayerInfo act_info)
95 {
96     ARM_COMPUTE_ERROR_ON_NULLPTR(src);
97 
98     auto padding_info = get_padding_info({ src, dst });
99 
100     _run_in_place = (dst == nullptr) || (dst == src);
101 
102     if(dst != nullptr)
103     {
104         // Destination auto inizialitation if not yet initialized
105         auto_init_if_empty(*dst, *src->clone());
106     }
107 
108     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, (dst != nullptr) ? dst : nullptr, act_info));
109 
110     const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / src->element_size(), src->dimension(0));
111 
112     const DataType dt      = src->data_type();
113     float          a_const = act_info.a();
114     float          b_const = act_info.b();
115 
116     const ActivationLayerInfo::ActivationFunction f_act        = act_info.activation();
117     const bool                                    is_quantized = is_data_type_quantized(dt);
118     const bool                                    perform_activation_in_float =
119         (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
120         || (f_act == ActivationLayerInfo::ActivationFunction::TANH)
121         || (f_act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
122         || (f_act == ActivationLayerInfo::ActivationFunction::LEAKY_RELU);
123 
124     // Set build options
125     CLBuildOptions build_opts;
126     build_opts.add_option_if(perform_activation_in_float, "-DFLOAT_DOMAIN");
127     build_opts.add_option_if(_run_in_place, "-DIN_PLACE");
128     build_opts.add_option("-DACT=" + lower_string(string_from_activation_func(f_act)));
129     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(dt));
130     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
131     build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % num_elems_processed_per_iteration));
132 
133     std::string kernel_name = std::string("activation_layer");
134 
135     // Set quantization info build options
136     if(is_quantized)
137     {
138         const UniformQuantizationInfo iq_info = src->quantization_info().uniform();
139 
140         if(!perform_activation_in_float)
141         {
142             int a_const_int = 0;
143             int b_const_int = 0;
144 
145             // Create quantized version of constants a, b if needed
146             switch(dt)
147             {
148                 case DataType::QASYMM8:
149                 {
150                     a_const_int = quantize_qasymm8(a_const, iq_info);
151                     b_const_int = quantize_qasymm8(b_const, iq_info);
152                 }
153                 break;
154                 case DataType::QASYMM8_SIGNED:
155                 {
156                     a_const_int = quantize_qasymm8_signed(a_const, iq_info);
157                     b_const_int = quantize_qasymm8_signed(b_const, iq_info);
158                 }
159                 break;
160                 case DataType::QSYMM16:
161                 {
162                     a_const_int = quantize_qsymm16(a_const, iq_info);
163                     b_const_int = quantize_qsymm16(b_const, iq_info);
164                 }
165                 break;
166                 default:
167                     break;
168             }
169             build_opts.add_option(("-DA_VAL=" + support::cpp11::to_string(a_const_int)));
170             build_opts.add_option(("-DB_VAL=" + support::cpp11::to_string(b_const_int)));
171         }
172         else
173         {
174             build_opts.add_option(("-DA_VAL=" + float_to_string_with_full_precision(a_const)));
175             build_opts.add_option(("-DB_VAL=" + float_to_string_with_full_precision(b_const)));
176         }
177 
178         // Quantized value of 0 corresponds to the offset o1
179         build_opts.add_option(("-DCONST_0=" + (is_data_type_quantized_asymmetric(dt) ? support::cpp11::to_string(iq_info.offset) : "0")));
180         build_opts.add_option(("-DS1_VAL=" + float_to_string_with_full_precision(iq_info.scale)));
181         build_opts.add_option_if(is_data_type_quantized_asymmetric(dt), "-DO1_VAL=" + support::cpp11::to_string(iq_info.offset));
182 
183         // Set correct kernel name
184         kernel_name += perform_activation_in_float ? std::string("_quant_f32") : std::string("_quant");
185 
186         // Set scale and offset of the source and destination if they have different quantization info
187         if(dst != nullptr)
188         {
189             const UniformQuantizationInfo oq_info = dst->quantization_info().uniform();
190 
191             if(iq_info != oq_info)
192             {
193                 build_opts.add_option(("-DS2_VAL=" + float_to_string_with_full_precision(oq_info.scale)));
194                 build_opts.add_option_if(is_data_type_quantized_asymmetric(dt), "-DO2_VAL=" + support::cpp11::to_string(oq_info.offset));
195             }
196         }
197     }
198     else
199     {
200         // Set A, B constants in build options for float types
201         build_opts.add_option(("-DA_VAL=" + float_to_string_with_full_precision(a_const)));
202         build_opts.add_option(("-DB_VAL=" + float_to_string_with_full_precision(b_const)));
203     }
204 
205     // Create kernel
206     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
207 
208     // Configure kernel window
209     Window win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration));
210     ICLKernel::configure_internal(win);
211 
212     // Set config_id for enabling LWS tuning
213     _config_id = "activation_layer_";
214     _config_id += lower_string(string_from_data_type(dt));
215     _config_id += "_";
216     _config_id += support::cpp11::to_string(src->dimension(0));
217     _config_id += "_";
218     _config_id += support::cpp11::to_string(src->dimension(1));
219 
220     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
221 }
222 
validate(const ITensorInfo * src,const ITensorInfo * dst,const ActivationLayerInfo & act_info)223 Status ClActivationKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const ActivationLayerInfo &act_info)
224 {
225     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, act_info));
226     return Status{};
227 }
228 
run_op(ITensorPack & tensors,const Window & window,::cl::CommandQueue & queue)229 void ClActivationKernel::run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue)
230 {
231     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
232     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
233 
234     const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
235     auto       dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
236     ARM_COMPUTE_ERROR_ON(_run_in_place && src != dst);
237 
238     Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
239     Window slice     = collapsed.first_slice_window_3D();
240 
241     do
242     {
243         unsigned int idx = 0;
244         add_3D_tensor_argument(idx, src, slice);
245         if(!_run_in_place)
246         {
247             add_3D_tensor_argument(idx, dst, slice);
248         }
249         enqueue(queue, *this, slice, lws_hint());
250     }
251     while(collapsed.slide_window_slice_3D(slice));
252 }
253 } // namespace kernels
254 } // namespace opencl
255 } // namespace arm_compute
256