Georgios Pinitas | c0b6f76 | 2020-11-02 01:37:17 +0000 | [diff] [blame] | 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" |
Georgios Pinitas | 40f51a6 | 2020-11-21 03:04:18 +0000 | [diff] [blame^] | 28 | |
Georgios Pinitas | c0b6f76 | 2020-11-02 01:37:17 +0000 | [diff] [blame] | 29 | #include <set> |
Georgios Pinitas | 40f51a6 | 2020-11-21 03:04:18 +0000 | [diff] [blame^] | 30 | |
Georgios Pinitas | c0b6f76 | 2020-11-02 01:37:17 +0000 | [diff] [blame] | 31 | namespace arm_compute |
| 32 | { |
| 33 | namespace |
| 34 | { |
| 35 | GEMMLowpOutputStageInfo calculate_output_stage_metadata(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const ActivationLayerInfo &act) |
| 36 | { |
| 37 | // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() |
| 38 | // Extract and negate input and weights offset |
| 39 | const QuantizationInfo iqinfo = input->quantization_info(); |
| 40 | const QuantizationInfo wqinfo = weights->quantization_info(); |
| 41 | const QuantizationInfo oqinfo = (output->total_size() == 0) ? iqinfo : output->quantization_info(); |
| 42 | const UniformQuantizationInfo uoqinfo = oqinfo.uniform(); |
| 43 | const DataType data_type = input->data_type(); |
| 44 | // Merge activation with output stage |
| 45 | const std::set<ActivationLayerInfo::ActivationFunction> supported_acts = { ActivationLayerInfo::ActivationFunction::RELU, |
| 46 | ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, |
| 47 | ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU |
| 48 | }; |
| 49 | PixelValue type_min{}; |
| 50 | PixelValue type_max{}; |
| 51 | std::tie(type_min, type_max) = get_min_max(data_type); |
| 52 | int32_t min_activation = type_min.get<int32_t>(); |
| 53 | int32_t max_activation = type_max.get<int32_t>(); |
| 54 | if(supported_acts.count(act.activation()) != 0) |
| 55 | { |
| 56 | std::tie(min_activation, max_activation) = get_quantized_activation_min_max(act, data_type, uoqinfo); |
| 57 | } |
| 58 | GEMMLowpOutputStageInfo os_info; |
| 59 | os_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; |
| 60 | os_info.gemmlowp_offset = uoqinfo.offset; |
| 61 | os_info.gemmlowp_min_bound = min_activation; |
| 62 | os_info.gemmlowp_max_bound = max_activation; |
| 63 | os_info.is_quantized_per_channel = (weights->data_type() == DataType::QSYMM8_PER_CHANNEL); |
| 64 | quantization::calculate_quantized_multipliers(iqinfo, wqinfo, oqinfo, os_info); |
| 65 | return os_info; |
| 66 | } |
| 67 | AsmGemmInfo init_assembly_metadata(const Conv2dInfo &info, bool is_indirect) |
| 68 | { |
| 69 | AsmGemmInfo asm_info; |
| 70 | asm_info.method = is_indirect ? AsmConvMethod::Indirect : AsmConvMethod::Conv; |
| 71 | asm_info.ps_info = info.conv_info; |
| 72 | asm_info.activation_info = info.act_info; |
| 73 | asm_info.depth_output_gemm3d = true; |
| 74 | asm_info.reinterpret_input_as_3d = true; |
| 75 | asm_info.padding_top = info.conv_info.pad_top(); |
| 76 | asm_info.padding_left = info.conv_info.pad_left(); |
| 77 | asm_info.padding_value = 0.f; |
| 78 | asm_info.negated_offsets = false; |
| 79 | return asm_info; |
| 80 | } |
| 81 | } // namespace |
| 82 | |
| 83 | NEGEMMConv2d::NEGEMMConv2d(const std::shared_ptr<IMemoryManager> &memory_manager) |
| 84 | : _gemm_asm_func(memory_manager), _activation_func(), _weights_permute_func(), _original_weights(nullptr), _permuted_weights(), _is_prepared(false), _run_activation(false) |
| 85 | { |
| 86 | } |
| 87 | void NEGEMMConv2d::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const Conv2dInfo &info) |
| 88 | { |
| 89 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); |
| 90 | ARM_COMPUTE_ERROR_THROW_ON(NEGEMMConv2d::validate(input->info(), |
| 91 | weights->info(), |
| 92 | biases != nullptr ? biases->info() : nullptr, |
| 93 | output->info(), |
| 94 | info)); |
| 95 | _original_weights = weights; |
| 96 | _weights_permute_func.configure(weights, &_permuted_weights, PermutationVector{ 3, 0, 1, 2 }); |
| 97 | |
| 98 | // Configure assembly dispatch |
| 99 | AsmGemmInfo asm_info = init_assembly_metadata(info, false); |
| 100 | if(is_data_type_quantized(input->info()->data_type())) |
| 101 | { |
| 102 | asm_info.output_stage = calculate_output_stage_metadata(input->info(), weights->info(), output->info(), info.act_info); |
| 103 | } |
| 104 | _gemm_asm_func.configure(input, &_permuted_weights, biases, output, asm_info); |
| 105 | |
| 106 | // Configure activation |
| 107 | if(info.act_info.enabled() && !_gemm_asm_func.is_activation_supported(info.act_info)) |
| 108 | { |
| 109 | _activation_func.configure(output, nullptr, info.act_info); |
| 110 | _run_activation = true; |
| 111 | } |
| 112 | } |
| 113 | Status NEGEMMConv2d::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const Conv2dInfo &info) |
| 114 | { |
| 115 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); |
| 116 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::BFLOAT16, DataType::F16, DataType::F32); |
| 117 | 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); |
| 118 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights); |
| 119 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.num_groups > 1, "Grouping (num_groups != 1) is not supported on NEON"); |
| 120 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_layout() != DataLayout::NHWC, "Data layout supported is NHWC"); |
| 121 | const DataType data_type = input->data_type(); |
| 122 | const TensorShape i_shape = input->tensor_shape(); |
| 123 | const TensorShape w_shape = weights->tensor_shape(); |
| 124 | ARM_COMPUTE_RETURN_ERROR_ON(w_shape[0] != i_shape[0]); |
| 125 | ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); |
| 126 | // Validate biases |
| 127 | if(biases != nullptr) |
| 128 | { |
| 129 | if(is_data_type_quantized_asymmetric(data_type)) |
| 130 | { |
| 131 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); |
| 132 | } |
| 133 | else if(data_type == DataType::BFLOAT16) |
| 134 | { |
| 135 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F32); |
| 136 | } |
| 137 | else |
| 138 | { |
| 139 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); |
| 140 | } |
| 141 | ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3)); |
| 142 | ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); |
| 143 | } |
| 144 | |
| 145 | AsmGemmInfo asm_info = init_assembly_metadata(info, false); |
| 146 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMAssemblyDispatch::validate(input, weights, biases, output, asm_info)); |
| 147 | return Status{}; |
| 148 | } |
| 149 | void NEGEMMConv2d::run() |
| 150 | { |
| 151 | prepare(); |
| 152 | |
| 153 | _gemm_asm_func.run(); |
| 154 | if(_run_activation) |
| 155 | { |
| 156 | _activation_func.run(); |
| 157 | } |
| 158 | } |
| 159 | void NEGEMMConv2d::prepare() |
| 160 | { |
| 161 | if(!_is_prepared) |
| 162 | { |
| 163 | _permuted_weights.allocator()->allocate(); |
| 164 | _weights_permute_func.run(); |
| 165 | _original_weights->mark_as_unused(); |
| 166 | _is_prepared = true; |
| 167 | } |
| 168 | } |
| 169 | } // namespace arm_compute |