1 /*
2 * Copyright (c) 2017-2020 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 "arm_compute/runtime/CL/functions/CLSoftmaxLayer.h"
25
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/Types.h"
29 #include "arm_compute/core/Utils.h"
30 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
31 #include "arm_compute/runtime/CL/CLScheduler.h"
32 #include "src/core/CL/ICLKernel.h"
33 #include "src/core/CL/kernels/CLFillBorderKernel.h"
34 #include "src/core/CL/kernels/CLSoftmaxLayerKernel.h"
35 #include "src/core/helpers/SoftmaxHelpers.h"
36 #include "support/MemorySupport.h"
37
38 namespace arm_compute
39 {
40 template <bool IS_LOG>
CLSoftmaxLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager)41 CLSoftmaxLayerGeneric<IS_LOG>::CLSoftmaxLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager)
42 : _memory_group(std::move(memory_manager)),
43 _permute_input(),
44 _permute_output(),
45 _max_shift_exp_sum_kernel(support::cpp14::make_unique<CLLogits1DMaxShiftExpSumKernel>()),
46 _norm_kernel(support::cpp14::make_unique<CLLogits1DNormKernel>()),
47 _max(),
48 _sum(),
49 _tmp(),
50 _input_permuted(),
51 _output_permuted(),
52 _needs_permute()
53 {
54 }
55
56 template <bool IS_LOG>
57 CLSoftmaxLayerGeneric<IS_LOG>::~CLSoftmaxLayerGeneric() = default;
58
59 template <bool IS_LOG>
configure(const ICLTensor * input,ICLTensor * output,float beta,int32_t axis)60 void CLSoftmaxLayerGeneric<IS_LOG>::configure(const ICLTensor *input, ICLTensor *output, float beta, int32_t axis)
61 {
62 configure(CLKernelLibrary::get().get_compile_context(), input, output, beta, axis);
63 }
64
65 template <bool IS_LOG>
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,float beta,int32_t axis)66 void CLSoftmaxLayerGeneric<IS_LOG>::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, float beta, int32_t axis)
67 {
68 // Perform validation step
69 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
70 ARM_COMPUTE_ERROR_THROW_ON(CLSoftmaxLayerGeneric<IS_LOG>::validate(input->info(), output->info(), beta, axis));
71
72 const size_t actual_axis = static_cast<size_t>(wrap_around(axis, static_cast<int32_t>(input->info()->num_dimensions())));
73
74 _needs_permute = actual_axis != 0;
75 ICLTensor *tmp_output = output;
76 const ICLTensor *tmp_input = _needs_permute ? &_input_permuted : input;
77 if(_needs_permute)
78 {
79 _memory_group.manage(&_input_permuted);
80 _memory_group.manage(&_output_permuted);
81 _permute_input.configure(compile_context, input, &_input_permuted, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
82 tmp_output = &_output_permuted;
83 }
84
85 // Create intermediate tensors
86 DataType tmp_data_type = is_data_type_quantized_asymmetric(tmp_input->info()->data_type()) ? DataType::S32 : tmp_input->info()->data_type();
87 TensorInfo tensor_info_tmp(tmp_input->info()->clone()->set_data_type(tmp_data_type));
88 _tmp.allocator()->init(tensor_info_tmp);
89 TensorShape max_sum_shape = tmp_input->info()->tensor_shape();
90 max_sum_shape.set(0, 1);
91 _max.allocator()->init(tmp_input->info()->clone()->set_tensor_shape(max_sum_shape));
92 _sum.allocator()->init(tmp_input->info()->clone()->set_tensor_shape(max_sum_shape).set_data_type(tmp_data_type));
93
94 // Set GPU target to kernels
95 _max_shift_exp_sum_kernel->set_target(CLScheduler::get().target());
96
97 // Manage intermediate buffers
98 _memory_group.manage(&_tmp);
99 _memory_group.manage(&_max);
100 _memory_group.manage(&_sum);
101
102 SoftmaxKernelInfo softmax_info;
103 softmax_info.beta = beta;
104 softmax_info.is_log = IS_LOG;
105 softmax_info.input_data_type = tmp_input->info()->data_type();
106
107 // Configure kernels
108 _max_shift_exp_sum_kernel->configure(compile_context, tmp_input, &_max, &_tmp, &_sum, softmax_info);
109 _norm_kernel->configure(compile_context, &_tmp, &_sum, tmp_output, softmax_info);
110
111 // Allocate intermediate buffers
112 _tmp.allocator()->allocate();
113 _max.allocator()->allocate();
114 _sum.allocator()->allocate();
115 if(_needs_permute)
116 {
117 _permute_output.configure(compile_context, &_output_permuted, output, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
118 _input_permuted.allocator()->allocate();
119 _output_permuted.allocator()->allocate();
120 }
121 }
122
123 template <bool IS_LOG>
validate(const ITensorInfo * input,const ITensorInfo * output,float beta,int32_t axis)124 Status CLSoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const ITensorInfo *output, float beta, int32_t axis)
125 {
126 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
127 ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 4, "Only up to 4 dimensions are supported");
128 ARM_COMPUTE_UNUSED(beta);
129 ARM_COMPUTE_RETURN_ERROR_ON(axis < static_cast<int32_t>(-input->num_dimensions()) || static_cast<int32_t>(input->num_dimensions()) <= axis);
130
131 const size_t actual_axis = static_cast<size_t>(wrap_around(axis, static_cast<int32_t>(input->num_dimensions())));
132 const bool needs_permute = actual_axis != 0;
133 if(needs_permute)
134 {
135 const PermutationVector permutation_vector = softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis);
136 const TensorShape permuted_shape = misc::shape_calculator::compute_permutation_output_shape(*input, permutation_vector);
137 TensorInfo input_permuted(input->clone()->set_tensor_shape(permuted_shape));
138 ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(input, &input_permuted, permutation_vector));
139 TensorInfo output_permuted(output->clone()->set_tensor_shape(permuted_shape));
140 ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(&output_permuted, output, permutation_vector));
141 }
142
143 // Create intermediate tensor info
144 DataType tmp_data_type = is_data_type_quantized_asymmetric(input->data_type()) ? DataType::S32 : input->data_type();
145 TensorInfo tensor_info_tmp(input->clone()->set_data_type(tmp_data_type).set_is_resizable(true));
146
147 TensorShape max_sum_shape = input->tensor_shape();
148 max_sum_shape.set(0, 1);
149 TensorInfo tensor_info_max(input->clone()->set_tensor_shape(max_sum_shape).set_is_resizable(true));
150 TensorInfo tensor_info_sum(input->clone()->set_tensor_shape(max_sum_shape).set_data_type(tmp_data_type).set_quantization_info(QuantizationInfo()).set_is_resizable(true));
151
152 SoftmaxKernelInfo softmax_info;
153 softmax_info.beta = beta;
154 softmax_info.is_log = IS_LOG;
155 softmax_info.input_data_type = input->data_type();
156
157 ARM_COMPUTE_RETURN_ON_ERROR(CLLogits1DMaxShiftExpSumKernel::validate(input, &tensor_info_max, &tensor_info_tmp, &tensor_info_sum));
158 ARM_COMPUTE_RETURN_ON_ERROR(CLLogits1DNormKernel::validate(&tensor_info_tmp, &tensor_info_sum, output, softmax_info));
159
160 return Status{};
161 }
162
163 template <bool IS_LOG>
run()164 void CLSoftmaxLayerGeneric<IS_LOG>::run()
165 {
166 MemoryGroupResourceScope scope_mg(_memory_group);
167
168 if(_needs_permute)
169 {
170 _permute_input.run();
171 }
172
173 CLScheduler::get().enqueue(*_max_shift_exp_sum_kernel, false);
174 CLScheduler::get().enqueue(*_norm_kernel, !_needs_permute);
175
176 if(_needs_permute)
177 {
178 _permute_output.run();
179 }
180 }
181
182 template class CLSoftmaxLayerGeneric<false>;
183 template class CLSoftmaxLayerGeneric<true>;
184
185 } // namespace arm_compute
186