• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2016-2020 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/core/CL/kernels/CLActivationLayerKernel.h"
25 
26 #include "arm_compute/core/CL/CLCoreRuntimeContext.h"
27 #include "arm_compute/core/CL/CLHelpers.h"
28 #include "arm_compute/core/CL/ICLTensor.h"
29 #include "arm_compute/core/TensorInfo.h"
30 #include "arm_compute/core/Utils.h"
31 #include "src/core/CL/CLValidate.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33 #include "src/core/helpers/WindowHelpers.h"
34 #include "support/Cast.h"
35 
36 #include "support/StringSupport.h"
37 
38 #include <set>
39 
40 namespace arm_compute
41 {
42 namespace
43 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,const ActivationLayerInfo & act_info)44 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info)
45 {
46     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
47     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16, DataType::F16, DataType::F32);
48 
49     static std::set<ActivationLayerInfo::ActivationFunction> quantized_supported_activations =
50     {
51         ActivationLayerInfo::ActivationFunction::RELU,
52         ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU,
53         ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
54         ActivationLayerInfo::ActivationFunction::LOGISTIC,
55         ActivationLayerInfo::ActivationFunction::TANH,
56         ActivationLayerInfo::ActivationFunction::HARD_SWISH
57     };
58     const DataType                                data_type = input->data_type();
59     const QuantizationInfo                       &oq_info   = (output != nullptr) ? output->quantization_info() : input->quantization_info();
60     const ActivationLayerInfo::ActivationFunction f_act     = act_info.activation();
61 
62     ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_quantized(data_type) && (quantized_supported_activations.count(f_act) == 0),
63                                     "For Quantized data type only tanh, logistic, relu and lower/upper bounded relu are supported");
64 
65     ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8 && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 128.f, 128)));
66     ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8 && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 256.f, 0)));
67 
68     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)));
69     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)));
70 
71     ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8_SIGNED && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 128.f, 0)));
72     ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8_SIGNED && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 256.f, -128)));
73 
74     // Checks performed when output is configured
75     if((output != nullptr) && (output->total_size() != 0))
76     {
77         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
78         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
79     }
80 
81     return Status{};
82 }
83 } // namespace
84 
CLActivationLayerKernel()85 CLActivationLayerKernel::CLActivationLayerKernel()
86     : _run_in_place(false)
87 {
88 }
89 
configure(const CLCompileContext & compile_context,ITensorInfo * input,ITensorInfo * output,ActivationLayerInfo act_info)90 void CLActivationLayerKernel::configure(const CLCompileContext &compile_context, ITensorInfo *input, ITensorInfo *output, ActivationLayerInfo act_info)
91 {
92     ARM_COMPUTE_ERROR_ON_NULLPTR(input);
93 
94     auto padding_info = get_padding_info({ input, output });
95 
96     _run_in_place = (output == nullptr) || (output == input);
97 
98     if(output != nullptr)
99     {
100         // Output auto inizialitation if not yet initialized
101         auto_init_if_empty(*output, *input->clone());
102     }
103 
104     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input, (output != nullptr) ? output : nullptr, act_info));
105 
106     const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / input->element_size(), input->dimension(0));
107 
108     const DataType dt      = input->data_type();
109     float          a_const = act_info.a();
110     float          b_const = act_info.b();
111 
112     const ActivationLayerInfo::ActivationFunction f_act        = act_info.activation();
113     const bool                                    is_quantized = is_data_type_quantized(dt);
114     const bool                                    perform_activation_in_float =
115         (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) || (f_act == ActivationLayerInfo::ActivationFunction::TANH) || (f_act == ActivationLayerInfo::ActivationFunction::HARD_SWISH);
116 
117     // Set build options
118     CLBuildOptions build_opts;
119     build_opts.add_option_if(perform_activation_in_float, "-DFLOAT_DOMAIN");
120     build_opts.add_option_if(_run_in_place, "-DIN_PLACE");
121     build_opts.add_option("-DACT=" + lower_string(string_from_activation_func(f_act)));
122     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(dt));
123     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
124     build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(input->dimension(0) % num_elems_processed_per_iteration));
125 
126     std::string kernel_name = std::string("activation_layer");
127 
128     // Set quantization info build options
129     if(is_quantized)
130     {
131         const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
132 
133         if(!perform_activation_in_float)
134         {
135             int a_const_int = 0;
136             int b_const_int = 0;
137 
138             // Create quantized version of constants a, b if needed
139             switch(dt)
140             {
141                 case DataType::QASYMM8:
142                 {
143                     a_const_int = quantize_qasymm8(a_const, iq_info);
144                     b_const_int = quantize_qasymm8(b_const, iq_info);
145                 }
146                 break;
147                 case DataType::QASYMM8_SIGNED:
148                 {
149                     a_const_int = quantize_qasymm8_signed(a_const, iq_info);
150                     b_const_int = quantize_qasymm8_signed(b_const, iq_info);
151                 }
152                 break;
153                 case DataType::QSYMM16:
154                 {
155                     a_const_int = quantize_qsymm16(a_const, iq_info);
156                     b_const_int = quantize_qsymm16(b_const, iq_info);
157                 }
158                 break;
159                 default:
160                     break;
161             }
162             build_opts.add_option(("-DA_VAL=" + support::cpp11::to_string(a_const_int)));
163             build_opts.add_option(("-DB_VAL=" + support::cpp11::to_string(b_const_int)));
164         }
165         else
166         {
167             build_opts.add_option(("-DA_VAL=" + float_to_string_with_full_precision(a_const)));
168             build_opts.add_option(("-DB_VAL=" + float_to_string_with_full_precision(b_const)));
169         }
170 
171         // Quantized value of 0 corresponds to the offset o1
172         build_opts.add_option(("-DCONST_0=" + (is_data_type_quantized_asymmetric(dt) ? support::cpp11::to_string(iq_info.offset) : "0")));
173         build_opts.add_option(("-DS1_VAL=" + float_to_string_with_full_precision(iq_info.scale)));
174         build_opts.add_option_if(is_data_type_quantized_asymmetric(dt), "-DO1_VAL=" + support::cpp11::to_string(iq_info.offset));
175 
176         // Set correct kernel name
177         kernel_name += perform_activation_in_float ? std::string("_quant_f32") : std::string("_quant");
178 
179         // Set scale and offset of the input and output if they have different quantization info
180         if(output != nullptr)
181         {
182             const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
183 
184             if(iq_info != oq_info)
185             {
186                 build_opts.add_option(("-DS2_VAL=" + float_to_string_with_full_precision(oq_info.scale)));
187                 build_opts.add_option_if(is_data_type_quantized_asymmetric(dt), "-DO2_VAL=" + support::cpp11::to_string(oq_info.offset));
188             }
189         }
190     }
191     else
192     {
193         // Set A, B constants in build options for float types
194         build_opts.add_option(("-DA_VAL=" + float_to_string_with_full_precision(a_const)));
195         build_opts.add_option(("-DB_VAL=" + float_to_string_with_full_precision(b_const)));
196     }
197 
198     // Create kernel
199     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
200 
201     // Configure kernel window
202     Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
203     ICLKernel::configure_internal(win);
204 
205     // Set config_id for enabling LWS tuning
206     _config_id = "activation_layer_";
207     _config_id += lower_string(string_from_data_type(dt));
208     _config_id += "_";
209     _config_id += support::cpp11::to_string(input->dimension(0));
210     _config_id += "_";
211     _config_id += support::cpp11::to_string(input->dimension(1));
212 
213     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
214 }
215 
validate(const ITensorInfo * input,const ITensorInfo * output,const ActivationLayerInfo & act_info)216 Status CLActivationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info)
217 {
218     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, act_info));
219     return Status{};
220 }
221 
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)222 void CLActivationLayerKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
223 {
224     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
225     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
226 
227     const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
228     auto       dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
229     ARM_COMPUTE_ERROR_ON(_run_in_place && src != dst);
230 
231     Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
232     Window slice     = collapsed.first_slice_window_3D();
233 
234     do
235     {
236         unsigned int idx = 0;
237         add_3D_tensor_argument(idx, src, slice);
238         if(!_run_in_place)
239         {
240             add_3D_tensor_argument(idx, dst, slice);
241         }
242         enqueue(queue, *this, slice, lws_hint());
243     }
244     while(collapsed.slide_window_slice_3D(slice));
245 }
246 } // namespace arm_compute
247