|
| 1 | +/* |
| 2 | + * Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved. |
| 3 | + * |
| 4 | + * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | + * you may not use this file except in compliance with the License. |
| 6 | + * You may obtain a copy of the License at |
| 7 | + * |
| 8 | + * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | + * |
| 10 | + * Unless required by applicable law or agreed to in writing, software |
| 11 | + * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | + * See the License for the specific language governing permissions and |
| 14 | + * limitations under the License. |
| 15 | + */ |
| 16 | + |
| 17 | +#ifndef ENTROPY_CALIBRATOR_HPP |
| 18 | +#define ENTROPY_CALIBRATOR_HPP |
| 19 | + |
| 20 | +#include <algorithm> |
| 21 | +#include <numeric> |
| 22 | +#include <iterator> |
| 23 | +#include "NvInfer.h" |
| 24 | + |
| 25 | +//! \class EntropyCalibratorImpl |
| 26 | +//! |
| 27 | +//! \brief Implements common functionality for Entropy calibrators. |
| 28 | +//! |
| 29 | +template <typename TBatchStream> |
| 30 | +class EntropyCalibratorImpl |
| 31 | +{ |
| 32 | +public: |
| 33 | + EntropyCalibratorImpl( |
| 34 | + TBatchStream stream, int firstBatch, std::string cal_table_name, const char* inputBlobName, bool readCache = true) |
| 35 | + : mStream{stream} |
| 36 | + , mCalibrationTableName(cal_table_name) |
| 37 | + , mInputBlobName(inputBlobName) |
| 38 | + , mReadCache(readCache) |
| 39 | + { |
| 40 | + nvinfer1::Dims4 dims = mStream.getDims(); |
| 41 | + mInputCount = std::accumulate( |
| 42 | + dims.d, dims.d + dims.nbDims, 1, std::multiplies<int64_t>()); |
| 43 | + cout << "dims.nbDims: " << dims.nbDims << endl; |
| 44 | + for (int i{0}; i < dims.nbDims; ++i) { |
| 45 | + cout << dims.d[i] << ", "; |
| 46 | + } |
| 47 | + cout << endl; |
| 48 | + |
| 49 | + cudaError_t state; |
| 50 | + state = cudaMalloc(&mDeviceInput, mInputCount * sizeof(float)); |
| 51 | + if (state) { |
| 52 | + cout << "allocate memory failed\n"; |
| 53 | + std::abort(); |
| 54 | + } |
| 55 | + cout << "mInputCount: " << mInputCount << endl; |
| 56 | + mStream.reset(firstBatch); |
| 57 | + } |
| 58 | + |
| 59 | + virtual ~EntropyCalibratorImpl() |
| 60 | + { |
| 61 | + cudaError_t state; |
| 62 | + state = cudaFree(mDeviceInput); |
| 63 | + if (state) { |
| 64 | + cout << "free memory failed\n"; |
| 65 | + std::abort(); |
| 66 | + } |
| 67 | + } |
| 68 | + |
| 69 | + int getBatchSize() const |
| 70 | + { |
| 71 | + return mStream.getBatchSize(); |
| 72 | + } |
| 73 | + |
| 74 | + bool getBatch(void* bindings[], const char* names[], int nbBindings) |
| 75 | + { |
| 76 | + if (!mStream.next()) |
| 77 | + { |
| 78 | + return false; |
| 79 | + } |
| 80 | + cudaError_t state; |
| 81 | + state = cudaMemcpy(mDeviceInput, mStream.getBatch(), mInputCount * sizeof(float), cudaMemcpyHostToDevice); |
| 82 | + if (state) { |
| 83 | + cout << "memory copy to device failed\n"; |
| 84 | + std::abort(); |
| 85 | + } |
| 86 | + assert(!strcmp(names[0], mInputBlobName)); |
| 87 | + bindings[0] = mDeviceInput; |
| 88 | + return true; |
| 89 | + } |
| 90 | + |
| 91 | + const void* readCalibrationCache(size_t& length) |
| 92 | + { |
| 93 | + mCalibrationCache.clear(); |
| 94 | + std::ifstream input(mCalibrationTableName, std::ios::binary); |
| 95 | + input >> std::noskipws; |
| 96 | + if (mReadCache && input.good()) |
| 97 | + { |
| 98 | + std::copy(std::istream_iterator<char>(input), std::istream_iterator<char>(), |
| 99 | + std::back_inserter(mCalibrationCache)); |
| 100 | + } |
| 101 | + length = mCalibrationCache.size(); |
| 102 | + return length ? mCalibrationCache.data() : nullptr; |
| 103 | + } |
| 104 | + |
| 105 | + void writeCalibrationCache(const void* cache, size_t length) |
| 106 | + { |
| 107 | + std::ofstream output(mCalibrationTableName, std::ios::binary); |
| 108 | + output.write(reinterpret_cast<const char*>(cache), length); |
| 109 | + } |
| 110 | + |
| 111 | +private: |
| 112 | + TBatchStream mStream; |
| 113 | + size_t mInputCount; |
| 114 | + std::string mCalibrationTableName; |
| 115 | + const char* mInputBlobName; |
| 116 | + bool mReadCache{true}; |
| 117 | + void* mDeviceInput{nullptr}; |
| 118 | + std::vector<char> mCalibrationCache; |
| 119 | +}; |
| 120 | + |
| 121 | +//! \class Int8EntropyCalibrator2 |
| 122 | +//! |
| 123 | +//! \brief Implements Entropy calibrator 2. |
| 124 | +//! CalibrationAlgoType is kENTROPY_CALIBRATION_2. |
| 125 | +//! |
| 126 | +template <typename TBatchStream> |
| 127 | +class Int8EntropyCalibrator2 : public nvinfer1::IInt8EntropyCalibrator2 |
| 128 | +{ |
| 129 | +public: |
| 130 | + Int8EntropyCalibrator2( |
| 131 | + TBatchStream stream, int firstBatch, const char* networkName, const char* inputBlobName, bool readCache = true) |
| 132 | + : mImpl(stream, firstBatch, networkName, inputBlobName, readCache) |
| 133 | + { |
| 134 | + } |
| 135 | + |
| 136 | + int getBatchSize() const noexcept override |
| 137 | + { |
| 138 | + return mImpl.getBatchSize(); |
| 139 | + } |
| 140 | + |
| 141 | + bool getBatch(void* bindings[], const char* names[], int nbBindings) noexcept override |
| 142 | + { |
| 143 | + return mImpl.getBatch(bindings, names, nbBindings); |
| 144 | + } |
| 145 | + |
| 146 | + const void* readCalibrationCache(size_t& length) noexcept override |
| 147 | + { |
| 148 | + return mImpl.readCalibrationCache(length); |
| 149 | + } |
| 150 | + |
| 151 | + void writeCalibrationCache(const void* cache, size_t length) noexcept override |
| 152 | + { |
| 153 | + mImpl.writeCalibrationCache(cache, length); |
| 154 | + } |
| 155 | + |
| 156 | +private: |
| 157 | + EntropyCalibratorImpl<TBatchStream> mImpl; |
| 158 | +}; |
| 159 | + |
| 160 | +#endif // ENTROPY_CALIBRATOR_H |
0 commit comments