1 /*
2 * Copyright (c) 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/NEON/functions/NEGEMMConv2d.h"
25 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
26 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
27 #include "arm_compute/runtime/NEON/NEScheduler.h"
28 #include <set>
29 namespace arm_compute
30 {
31 namespace
32 {
calculate_output_stage_metadata(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * output,const ActivationLayerInfo & act)33 GEMMLowpOutputStageInfo calculate_output_stage_metadata(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const ActivationLayerInfo &act)
34 {
35 // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
36 // Extract and negate input and weights offset
37 const QuantizationInfo iqinfo = input->quantization_info();
38 const QuantizationInfo wqinfo = weights->quantization_info();
39 const QuantizationInfo oqinfo = (output->total_size() == 0) ? iqinfo : output->quantization_info();
40 const UniformQuantizationInfo uoqinfo = oqinfo.uniform();
41 const DataType data_type = input->data_type();
42 // Merge activation with output stage
43 const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU,
44 ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
45 ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU
46 };
47 PixelValue type_min{};
48 PixelValue type_max{};
49 std::tie(type_min, type_max) = get_min_max(data_type);
50 int32_t min_activation = type_min.get<int32_t>();
51 int32_t max_activation = type_max.get<int32_t>();
52 if(supported_acts.count(act.activation()) != 0)
53 {
54 std::tie(min_activation, max_activation) = get_quantized_activation_min_max(act, data_type, uoqinfo);
55 }
56 GEMMLowpOutputStageInfo os_info;
57 os_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
58 os_info.gemmlowp_offset = uoqinfo.offset;
59 os_info.gemmlowp_min_bound = min_activation;
60 os_info.gemmlowp_max_bound = max_activation;
61 os_info.is_quantized_per_channel = (weights->data_type() == DataType::QSYMM8_PER_CHANNEL);
62 quantization::calculate_quantized_multipliers(iqinfo, wqinfo, oqinfo, os_info);
63 return os_info;
64 }
init_assembly_metadata(const Conv2dInfo & info,bool is_indirect)65 AsmGemmInfo init_assembly_metadata(const Conv2dInfo &info, bool is_indirect)
66 {
67 AsmGemmInfo asm_info;
68 asm_info.method = is_indirect ? AsmConvMethod::Indirect : AsmConvMethod::Conv;
69 asm_info.ps_info = info.conv_info;
70 asm_info.activation_info = info.act_info;
71 asm_info.depth_output_gemm3d = true;
72 asm_info.reinterpret_input_as_3d = true;
73 asm_info.padding_top = info.conv_info.pad_top();
74 asm_info.padding_left = info.conv_info.pad_left();
75 asm_info.padding_value = 0.f;
76 asm_info.negated_offsets = false;
77 return asm_info;
78 }
79 } // namespace
80
NEGEMMConv2d(const std::shared_ptr<IMemoryManager> & memory_manager)81 NEGEMMConv2d::NEGEMMConv2d(const std::shared_ptr<IMemoryManager> &memory_manager)
82 : _gemm_asm_func(memory_manager), _activation_func(), _weights_permute_func(), _original_weights(nullptr), _permuted_weights(), _is_prepared(false), _run_activation(false)
83 {
84 }
configure(ITensor * input,const ITensor * weights,const ITensor * biases,ITensor * output,const Conv2dInfo & info)85 void NEGEMMConv2d::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const Conv2dInfo &info)
86 {
87 ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
88 ARM_COMPUTE_ERROR_THROW_ON(NEGEMMConv2d::validate(input->info(),
89 weights->info(),
90 biases != nullptr ? biases->info() : nullptr,
91 output->info(),
92 info));
93 _original_weights = weights;
94 _weights_permute_func.configure(weights, &_permuted_weights, PermutationVector{ 3, 0, 1, 2 });
95
96 // Configure assembly dispatch
97 AsmGemmInfo asm_info = init_assembly_metadata(info, false);
98 if(is_data_type_quantized(input->info()->data_type()))
99 {
100 asm_info.output_stage = calculate_output_stage_metadata(input->info(), weights->info(), output->info(), info.act_info);
101 }
102 _gemm_asm_func.configure(input, &_permuted_weights, biases, output, asm_info);
103
104 // Configure activation
105 if(info.act_info.enabled() && !_gemm_asm_func.is_activation_supported(info.act_info))
106 {
107 _activation_func.configure(output, nullptr, info.act_info);
108 _run_activation = true;
109 }
110 }
validate(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * output,const Conv2dInfo & info)111 Status NEGEMMConv2d::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const Conv2dInfo &info)
112 {
113 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
114 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::BFLOAT16, DataType::F16, DataType::F32);
115 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL, DataType::BFLOAT16, DataType::F16, DataType::F32);
116 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights);
117 ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.num_groups > 1, "Grouping (num_groups != 1) is not supported on NEON");
118 ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_layout() != DataLayout::NHWC, "Data layout supported is NHWC");
119 const DataType data_type = input->data_type();
120 const TensorShape i_shape = input->tensor_shape();
121 const TensorShape w_shape = weights->tensor_shape();
122 ARM_COMPUTE_RETURN_ERROR_ON(w_shape[0] != i_shape[0]);
123 ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
124 // Validate biases
125 if(biases != nullptr)
126 {
127 if(is_data_type_quantized_asymmetric(data_type))
128 {
129 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
130 }
131 else if(data_type == DataType::BFLOAT16)
132 {
133 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F32);
134 }
135 else
136 {
137 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
138 }
139 ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3));
140 ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
141 }
142
143 AsmGemmInfo asm_info = init_assembly_metadata(info, false);
144 ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMAssemblyDispatch::validate(input, weights, biases, output, asm_info));
145 return Status{};
146 }
run()147 void NEGEMMConv2d::run()
148 {
149 prepare();
150
151 _gemm_asm_func.run();
152 if(_run_activation)
153 {
154 _activation_func.run();
155 }
156 }
prepare()157 void NEGEMMConv2d::prepare()
158 {
159 if(!_is_prepared)
160 {
161 _permuted_weights.allocator()->allocate();
162 _weights_permute_func.run();
163 _original_weights->mark_as_unused();
164 _is_prepared = true;
165 }
166 }
167 } // namespace arm_compute
168