xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/CpuActivationKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2022 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/cpu/kernels/CpuActivationKernel.h"
25 
26 #include "arm_compute/core/ITensor.h"
27 #include "arm_compute/core/TensorInfo.h"
28 #include "arm_compute/core/Utils.h"
29 #include "src/core/CPP/Validate.h"
30 #include "src/core/helpers/AutoConfiguration.h"
31 #include "src/core/helpers/WindowHelpers.h"
32 
33 #include "src/core/common/Registrars.h"
34 #include "src/cpu/kernels/activation/list.h"
35 
36 #include <array>
37 
38 namespace arm_compute
39 {
40 namespace cpu
41 {
42 namespace kernels
43 {
44 namespace
45 {
46 static const std::vector<CpuActivationKernel::ActivationKernel> available_kernels =
47 {
48 #ifdef ARM_COMPUTE_ENABLE_SVE
49     {
50         "sve_q8_activation_lut",
__anon86be97850202() 51         [](const ActivationDataTypeISASelectorData & data) { return ActivationLayerInfo::is_lut_supported(data.f, data.dt) && data.cpumodel == CPUModel::A510 && data.isa.sve; },
52         REGISTER_QASYMM8_SVE(arm_compute::cpu::sve_q8_activation_lut)
53     },
54 #endif // ARM_COMPUTE_ENABLE_SVE
55 #ifdef __aarch64__
56     {
57         // Neon LUT implementantion takes precedence
58         "neon_q8_activation_lut",
__anon86be97850302() 59         [](const ActivationDataTypeISASelectorData & data) { return ActivationLayerInfo::is_lut_supported(data.f, data.dt); },
60         REGISTER_Q8_NEON(arm_compute::cpu::neon_q8_activation_lut)
61     },
62 #endif // __aarch64__
63     {
64         "sve2_qu8_activation",
__anon86be97850402() 65         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8 && data.isa.sve2 && data.f != ActivationLayerInfo::ActivationFunction::GELU; },
66         REGISTER_QASYMM8_SVE2(arm_compute::cpu::sve2_qasymm8_activation)
67     },
68     {
69         "sve2_qs8_activation",
__anon86be97850502() 70         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED && data.isa.sve2 && data.f != ActivationLayerInfo::ActivationFunction::GELU; },
71         REGISTER_QASYMM8_SIGNED_SVE2(arm_compute::cpu::sve2_qasymm8_signed_activation)
72     },
73     {
74         "sve2_qs16_activation",
__anon86be97850602() 75         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QSYMM16 && data.isa.sve2 && data.f != ActivationLayerInfo::ActivationFunction::GELU; },
76         REGISTER_QSYMM16_SVE2(arm_compute::cpu::sve2_qsymm16_activation)
77     },
78     {
79         "sve_fp16_activation",
__anon86be97850702() 80         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::F16 && data.isa.sve && data.isa.fp16 && data.f != ActivationLayerInfo::ActivationFunction::GELU; },
81         REGISTER_FP16_SVE(arm_compute::cpu::sve_fp16_activation)
82     },
83     {
84         "sve_fp32_activation",
__anon86be97850802() 85         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::F32 && data.isa.sve && data.f != ActivationLayerInfo::ActivationFunction::GELU; },
86         REGISTER_FP32_SVE(arm_compute::cpu::sve_fp32_activation)
87     },
88     {
89         "neon_fp16_activation",
__anon86be97850902() 90         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::F16 && data.isa.fp16; },
91         REGISTER_FP16_NEON(arm_compute::cpu::neon_fp16_activation)
92     },
93     {
94         "neon_fp32_activation",
__anon86be97850a02() 95         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::F32; },
96         REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_activation)
97     },
98     {
99         "neon_qu8_activation",
__anon86be97850b02() 100         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8; },
101         REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_activation)
102     },
103     {
104         "neon_qs8_activation",
__anon86be97850c02() 105         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; },
106         REGISTER_QASYMM8_SIGNED_NEON(arm_compute::cpu::neon_qasymm8_signed_activation)
107     },
108     {
109         "neon_qs16_activation",
__anon86be97850d02() 110         [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QSYMM16; },
111         REGISTER_QSYMM16_NEON(arm_compute::cpu::neon_qsymm16_activation)
112     },
113 };
114 
115 /* Supported activation in the 8-bit integer domain */
116 static const std::array<ActivationLayerInfo::ActivationFunction, 8> qasymm8_activations =
117 {
118     ActivationLayerInfo::ActivationFunction::RELU,
119     ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU,
120     ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
121     ActivationLayerInfo::ActivationFunction::LOGISTIC,
122     ActivationLayerInfo::ActivationFunction::TANH,
123     ActivationLayerInfo::ActivationFunction::HARD_SWISH,
124     ActivationLayerInfo::ActivationFunction::LEAKY_RELU,
125     ActivationLayerInfo::ActivationFunction::GELU,
126 };
127 /* Supported activation in the 16-bit integer domain */
128 static const std::array<ActivationLayerInfo::ActivationFunction, 4> qsymm16_activations =
129 {
130     ActivationLayerInfo::ActivationFunction::LOGISTIC,
131     ActivationLayerInfo::ActivationFunction::TANH,
132     ActivationLayerInfo::ActivationFunction::HARD_SWISH,
133     ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU
134 };
135 
validate_arguments(const ITensorInfo * src,const ITensorInfo * dst,const ActivationLayerInfo & activation_info)136 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const ActivationLayerInfo &activation_info)
137 {
138     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src);
139     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8_SIGNED, DataType::QASYMM8, DataType::QSYMM16, DataType::F16, DataType::F32);
140 
141     const auto *uk = CpuActivationKernel::get_implementation(ActivationDataTypeISASelectorData{ src->data_type(), CPUInfo::get().get_cpu_model(), CPUInfo::get().get_isa(), activation_info.activation() });
142     ARM_COMPUTE_RETURN_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
143 
144     const DataType                                data_type = src->data_type();
145     const QuantizationInfo                       &oq_info   = (dst != nullptr) ? dst->quantization_info() : src->quantization_info();
146     const ActivationLayerInfo::ActivationFunction f_act     = activation_info.activation();
147 
148     ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_quantized_asymmetric(data_type) && (std::find(std::begin(qasymm8_activations), std::end(qasymm8_activations), f_act) == std::end(qasymm8_activations)),
149                                     "For QASYMM8 only hard swish, leaky relu, tanh, logistic, relu and lower/upper bounded relu are supported");
150 
151     ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_quantized_symmetric(data_type) && (std::find(std::begin(qsymm16_activations), std::end(qsymm16_activations), f_act) == std::end(qsymm16_activations)),
152                                     "For QSYMM16 only tanh and logistic are supported");
153     ARM_COMPUTE_RETURN_ERROR_ON((data_type == DataType::QASYMM8 || data_type == DataType::QASYMM16) && (f_act == ActivationLayerInfo::ActivationFunction::TANH)
154                                 && (oq_info != QuantizationInfo(1.f / 128.f, 128)));
155     ARM_COMPUTE_RETURN_ERROR_ON((data_type == DataType::QASYMM8 || data_type == DataType::QASYMM16) && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
156                                 && (oq_info != QuantizationInfo(1.f / 256.f, 0)));
157 
158     ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8_SIGNED && (f_act == ActivationLayerInfo::ActivationFunction::TANH) && (oq_info != QuantizationInfo(1.f / 128.f, 0)));
159     ARM_COMPUTE_RETURN_ERROR_ON(data_type == DataType::QASYMM8_SIGNED && (f_act == ActivationLayerInfo::ActivationFunction::LOGISTIC) && (oq_info != QuantizationInfo(1.f / 256.f, -128)));
160 
161     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)));
162     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)));
163 
164     // Checks performed when dst is configured
165     if((dst != nullptr) && (dst->total_size() != 0))
166     {
167         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst);
168         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
169     }
170 
171     return Status{};
172 }
173 
validate_and_configure_window(const ITensorInfo * src,ITensorInfo * dst)174 std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *src, ITensorInfo *dst)
175 {
176     // Configure kernel window
177     Window win = calculate_max_window(*src, Steps());
178 
179     if(dst != nullptr)
180     {
181         // dst auto inizialitation if not yet initialized
182         auto_init_if_empty(*dst, *src->clone());
183     }
184 
185     return std::make_pair(Status{}, win);
186 }
187 } // namespace
188 
configure(const ITensorInfo * src,ITensorInfo * dst,ActivationLayerInfo activation_info)189 void CpuActivationKernel::configure(const ITensorInfo *src, ITensorInfo *dst, ActivationLayerInfo activation_info)
190 {
191     ARM_COMPUTE_UNUSED(dst);
192     ARM_COMPUTE_ERROR_ON_NULLPTR(src);
193     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, activation_info));
194 
195     const auto uk = CpuActivationKernel::get_implementation(ActivationDataTypeISASelectorData{ src->data_type(), CPUInfo::get().get_cpu_model(), CPUInfo::get().get_isa(), activation_info.activation() });
196     if(dst != nullptr)
197     {
198         // dst auto inizialitation if not yet initialized
199         auto_init_if_empty(*dst, *src->clone());
200     }
201 
202     ARM_COMPUTE_ERROR_ON_NULLPTR(uk);
203 
204     _run_method = uk->ukernel;
205     _name       = std::string("CpuActivationKernel").append("/").append(uk->name);
206 
207 #ifdef __aarch64__
208     if(ActivationLayerInfo::is_lut_supported(activation_info.activation(), src->data_type()))
209     {
210         activation_info.init_lut(src->data_type(), src->quantization_info().uniform(), (dst) ? dst->quantization_info().uniform() : src->quantization_info().uniform());
211     }
212 #endif // __aarch64__
213     _act_info = activation_info;
214 
215     Window win;
216 
217     // Use squashed window
218     std::tie(win, _split_dimension) = calculate_squashed_or_max_window(*src);
219     ICPPKernel::configure(win);
220 }
221 
validate(const ITensorInfo * src,const ITensorInfo * dst,const ActivationLayerInfo & act_info)222 Status CpuActivationKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const ActivationLayerInfo &act_info)
223 {
224     ARM_COMPUTE_UNUSED(act_info);
225     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, act_info));
226     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), (dst != nullptr) ? dst->clone().get() : nullptr).first);
227 
228     return Status{};
229 }
230 
get_mws(const CPUInfo & platform,size_t thread_count) const231 size_t CpuActivationKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
232 {
233     ARM_COMPUTE_UNUSED(thread_count);
234     ARM_COMPUTE_UNUSED(platform);
235 
236     if(_split_dimension == Window::DimX)
237     {
238         // Don't split the work load too small if the tensor has been reinterpreted as 1D.
239         // This number is loosely chosen as threading overhead in each platform varies wildly.
240         return 1536;
241     }
242     return default_mws;
243 }
244 
run_op(ITensorPack & tensors,const Window & window,const ThreadInfo & info)245 void CpuActivationKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
246 {
247     // Early exit on disabled activation
248     if(!_act_info.enabled())
249     {
250         return;
251     }
252 
253     ARM_COMPUTE_UNUSED(info);
254     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
255     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
256 
257     ARM_COMPUTE_ERROR_ON(tensors.empty());
258     ARM_COMPUTE_ERROR_ON(_run_method == nullptr);
259 
260     const ITensor *src = tensors.get_const_tensor(TensorType::ACL_SRC);
261     ITensor       *dst = tensors.get_tensor(TensorType::ACL_DST);
262 
263     _run_method(src, dst, _act_info, window);
264 }
265 
name() const266 const char *CpuActivationKernel::name() const
267 {
268     return _name.c_str();
269 }
270 
get_available_kernels()271 const std::vector<CpuActivationKernel::ActivationKernel> &CpuActivationKernel::get_available_kernels()
272 {
273     return available_kernels;
274 }
275 } // namespace kernels
276 } // namespace cpu
277 } // namespace arm_compute
278