xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/operators/CpuSoftmax.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 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/cpu/operators/CpuSoftmax.h"
25 
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/TensorInfo.h"
28 #include "arm_compute/core/Validate.h"
29 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
30 #include "arm_compute/runtime/NEON/NEScheduler.h"
31 #include "src/common/utils/Log.h"
32 #include "src/core/helpers/MemoryHelpers.h"
33 #include "src/core/helpers/SoftmaxHelpers.h"
34 #include "src/cpu/kernels/CpuSoftmaxKernel.h"
35 #include "src/cpu/utils/CpuAuxTensorHandler.h"
36 
37 using namespace arm_compute::experimental;
38 
39 namespace arm_compute
40 {
41 namespace cpu
42 {
43 template <bool IS_LOG>
CpuSoftmaxGeneric()44 CpuSoftmaxGeneric<IS_LOG>::CpuSoftmaxGeneric()
45     : _permute_input(),
46       _permute_output(),
47       _max_kernel(),
48       _softmax_kernel(),
49       _max(),
50       _tmp(),
51       _input_permuted(),
52       _output_permuted(),
53       _needs_permute(false),
54       _aux_mem(InternalTensorIdx::COUNT)
55 {
56 }
57 
58 template <bool IS_LOG>
configure(const ITensorInfo * src,ITensorInfo * dst,float beta,int32_t axis)59 void CpuSoftmaxGeneric<IS_LOG>::configure(const ITensorInfo *src, ITensorInfo *dst, float beta, int32_t axis)
60 {
61     // Perform validation step
62     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
63     ARM_COMPUTE_ERROR_THROW_ON(CpuSoftmaxGeneric::validate(src, dst, beta, axis));
64     ARM_COMPUTE_LOG_PARAMS(src, dst, beta, axis);
65 
66     const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(src->num_dimensions())));
67 
68     _needs_permute = actual_axis > 0;
69 
70     if(_needs_permute)
71     {
72         _permute_input.configure(src, &_input_permuted, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
73     }
74 
75     // We want to deal with a 2D input. Either it is the permuted version of the original input (4D case)
76     // or it is the original input case (2D case)
77     const ITensorInfo *tmp_input = (_needs_permute ? &_input_permuted : src);
78 
79     // Create intermediate tensors shapes
80     TensorShape max_sum_shape = tmp_input->tensor_shape();
81     max_sum_shape.set(0, 1);
82     const TensorInfo input_info    = tmp_input->clone()->reset_padding().set_is_resizable(true);
83     DataType         tmp_data_type = is_data_type_quantized_asymmetric(tmp_input->data_type()) ? DataType::F32 : tmp_input->data_type();
84     TensorInfo       tensor_info_tmp(input_info.clone()->set_data_type(tmp_data_type));
85     TensorInfo       max_info(tmp_input->clone()->set_tensor_shape(max_sum_shape));
86 
87     // Init intermediate tensors
88     _max = TensorInfo(max_info);
89     _tmp = TensorInfo(tensor_info_tmp);
90 
91     // Configure kernels
92     auto mk = std::make_unique<kernels::CpuLogits1DMaxKernel>();
93     mk->configure(tmp_input, &_max);
94     _max_kernel = std::move(mk);
95 
96     auto sm = std::make_unique<kernels::CpuLogits1DSoftmaxKernel<IS_LOG>>();
97     if(_needs_permute)
98     {
99         // The normalization kernel stores the result in a permuted output tensor
100         sm->configure(tmp_input, &_max, &_output_permuted, beta, &_tmp);
101 
102         // Re-permute the permuted output into the requested (4D) output
103         _permute_output.configure(&_output_permuted, dst, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
104     }
105     else
106     {
107         // Softmax 2D case
108         sm->configure(tmp_input, &_max, dst, beta, &_tmp);
109     }
110     _softmax_kernel = std::move(sm);
111 
112     _aux_mem[InternalTensorIdx::MAX] = MemoryInfo(offset_int_vec(InternalTensorIdx::MAX), MemoryLifetime::Temporary, _max.total_size());
113     _aux_mem[InternalTensorIdx::TMP] = MemoryInfo(offset_int_vec(InternalTensorIdx::TMP), MemoryLifetime::Temporary, _tmp.total_size());
114 
115     _aux_mem[InternalTensorIdx::PERMUTED_SRC] = MemoryInfo(offset_int_vec(InternalTensorIdx::PERMUTED_SRC), MemoryLifetime::Temporary, _input_permuted.total_size());
116     _aux_mem[InternalTensorIdx::PERMUTED_DST] = MemoryInfo(offset_int_vec(InternalTensorIdx::PERMUTED_DST), MemoryLifetime::Temporary, _output_permuted.total_size());
117 }
118 
119 template <bool IS_LOG>
validate(const ITensorInfo * src,const ITensorInfo * dst,float beta,int32_t axis)120 Status CpuSoftmaxGeneric<IS_LOG>::validate(const ITensorInfo *src, const ITensorInfo *dst, float beta, int32_t axis)
121 {
122     // Perform validation step
123     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
124     ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->num_dimensions() > 4, "Only up to 4 dimensions are supported");
125     ARM_COMPUTE_UNUSED(beta);
126     ARM_COMPUTE_RETURN_ERROR_ON(axis < static_cast<int32_t>(-src->num_dimensions()) || static_cast<int32_t>(src->num_dimensions()) <= axis);
127 
128     // Create intermediate tensor info
129     DataType         tmp_data_type = src->data_type();
130     const TensorInfo tensor_info_tmp(src->clone()->set_data_type(tmp_data_type).set_is_resizable(true));
131 
132     TensorShape max_sum_shape = src->tensor_shape();
133     max_sum_shape.set(0, 1);
134     const TensorInfo tensor_info_max_sum(src->clone()->set_tensor_shape(max_sum_shape).set_data_type(tmp_data_type).set_quantization_info(src->quantization_info()).set_is_resizable(true));
135     const TensorInfo dont_care;
136 
137     const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(src->num_dimensions())));
138 
139     const bool needs_permute = actual_axis > 0;
140 
141     if(needs_permute)
142     {
143         const PermutationVector permutation_vector = softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis);
144         const TensorShape       permuted_shape     = misc::shape_calculator::compute_permutation_output_shape(*src, permutation_vector);
145         TensorInfo              input_permuted(src->clone()->set_tensor_shape(permuted_shape));
146         ARM_COMPUTE_RETURN_ON_ERROR(CpuPermute::validate(src, &input_permuted, permutation_vector));
147         TensorInfo output_permuted(dst->clone()->set_tensor_shape(permuted_shape));
148         ARM_COMPUTE_RETURN_ON_ERROR(CpuPermute::validate(&output_permuted, dst, permutation_vector));
149     }
150 
151     ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuLogits1DMaxKernel::validate(src, &tensor_info_max_sum));
152     ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuLogits1DSoftmaxKernel<IS_LOG>::validate(&tensor_info_tmp, &tensor_info_max_sum, dst, beta, &dont_care));
153 
154     return Status{};
155 }
156 
157 template <bool IS_LOG>
run(ITensorPack & tensors)158 void CpuSoftmaxGeneric<IS_LOG>::run(ITensorPack &tensors)
159 {
160     ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
161 
162     auto src = tensors.get_const_tensor(TensorType::ACL_SRC);
163     auto dst = tensors.get_tensor(TensorType::ACL_DST);
164 
165     CpuAuxTensorHandler tmp(offset_int_vec(InternalTensorIdx::TMP), _tmp, tensors, true);
166     CpuAuxTensorHandler max(offset_int_vec(InternalTensorIdx::MAX), _max, tensors, true);
167 
168     CpuAuxTensorHandler input_permuted(offset_int_vec(InternalTensorIdx::PERMUTED_SRC), _input_permuted, tensors, true);
169     CpuAuxTensorHandler output_permuted(offset_int_vec(InternalTensorIdx::PERMUTED_DST), _output_permuted, tensors, true);
170 
171     ITensorPack max_pack;
172     ITensorPack softmax_pack;
173 
174     if(_needs_permute)
175     {
176         ITensorPack permute_in_pack = { { TensorType::ACL_SRC, src }, { TensorType::ACL_DST, input_permuted.get() } };
177         _permute_input.run(permute_in_pack);
178 
179         max_pack = { { TensorType::ACL_SRC, input_permuted.get() }, { TensorType::ACL_DST, max.get() } };
180 
181         softmax_pack =
182         {
183             { TensorType::ACL_SRC_0, input_permuted.get() },
184             { TensorType::ACL_SRC_1, max.get() },
185             { TensorType::ACL_DST_0, output_permuted.get() },
186             { TensorType::ACL_DST_1, tmp.get() }
187         };
188     }
189     else
190     {
191         max_pack = { { TensorType::ACL_SRC, src }, { TensorType::ACL_DST, max.get() } };
192 
193         softmax_pack =
194         {
195             { TensorType::ACL_SRC_0, src },
196             { TensorType::ACL_SRC_1, max.get() },
197             { TensorType::ACL_DST_0, dst },
198             { TensorType::ACL_DST_1, tmp.get() }
199         };
200     }
201 
202     NEScheduler::get().schedule_op(_max_kernel.get(), Window::DimY, _max_kernel->window(), max_pack);
203     NEScheduler::get().schedule_op(_softmax_kernel.get(), Window::DimY, _softmax_kernel->window(), softmax_pack);
204 
205     if(_needs_permute)
206     {
207         ITensorPack permute_out_pack;
208         permute_out_pack.add_tensor(TensorType::ACL_SRC, output_permuted.get());
209         permute_out_pack.add_tensor(TensorType::ACL_DST, dst);
210         _permute_output.run(permute_out_pack);
211     }
212 }
213 
214 template <bool                   IS_LOG>
workspace() const215 experimental::MemoryRequirements CpuSoftmaxGeneric<IS_LOG>::workspace() const
216 {
217     return _aux_mem;
218 }
219 
220 template class CpuSoftmaxGeneric<false>;
221 template class CpuSoftmaxGeneric<true>;
222 } // namespace cpu
223 } // namespace arm_compute
224