1 /* 2 * Copyright (C) 2017 The Android Open Source Project 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 #include "quantization.h" 18 19 #include "util/base/logging.h" 20 21 namespace libtextclassifier2 { 22 namespace { 23 float DequantizeValue(int num_sparse_features, int quantization_bias, 24 float multiplier, int value) { 25 return 1.0 / num_sparse_features * (value - quantization_bias) * multiplier; 26 } 27 28 void DequantizeAdd8bit(const float* scales, const uint8* embeddings, 29 int bytes_per_embedding, const int num_sparse_features, 30 const int bucket_id, float* dest, int dest_size) { 31 static const int kQuantizationBias8bit = 128; 32 const float multiplier = scales[bucket_id]; 33 for (int k = 0; k < dest_size; ++k) { 34 dest[k] += 35 DequantizeValue(num_sparse_features, kQuantizationBias8bit, multiplier, 36 embeddings[bucket_id * bytes_per_embedding + k]); 37 } 38 } 39 40 void DequantizeAddNBit(const float* scales, const uint8* embeddings, 41 int bytes_per_embedding, int num_sparse_features, 42 int quantization_bits, int bucket_id, float* dest, 43 int dest_size) { 44 const int quantization_bias = 1 << (quantization_bits - 1); 45 const float multiplier = scales[bucket_id]; 46 for (int i = 0; i < dest_size; ++i) { 47 const int bit_offset = i * quantization_bits; 48 const int read16_offset = bit_offset / 8; 49 50 uint16 data = embeddings[bucket_id * bytes_per_embedding + read16_offset]; 51 // If we are not at the end of the embedding row, we can read 2-byte uint16, 52 // but if we are, we need to only read uint8. 53 if (read16_offset < bytes_per_embedding - 1) { 54 data |= embeddings[bucket_id * bytes_per_embedding + read16_offset + 1] 55 << 8; 56 } 57 int value = (data >> (bit_offset % 8)) & ((1 << quantization_bits) - 1); 58 dest[i] += DequantizeValue(num_sparse_features, quantization_bias, 59 multiplier, value); 60 } 61 } 62 } // namespace 63 64 bool CheckQuantizationParams(int bytes_per_embedding, int quantization_bits, 65 int output_embedding_size) { 66 if (bytes_per_embedding * 8 / quantization_bits < output_embedding_size) { 67 return false; 68 } 69 70 return true; 71 } 72 73 bool DequantizeAdd(const float* scales, const uint8* embeddings, 74 int bytes_per_embedding, int num_sparse_features, 75 int quantization_bits, int bucket_id, float* dest, 76 int dest_size) { 77 if (quantization_bits == 8) { 78 DequantizeAdd8bit(scales, embeddings, bytes_per_embedding, 79 num_sparse_features, bucket_id, dest, dest_size); 80 } else if (quantization_bits != 8) { 81 DequantizeAddNBit(scales, embeddings, bytes_per_embedding, 82 num_sparse_features, quantization_bits, bucket_id, dest, 83 dest_size); 84 } else { 85 TC_LOG(ERROR) << "Unsupported quantization_bits: " << quantization_bits; 86 return false; 87 } 88 89 return true; 90 } 91 92 } // namespace libtextclassifier2 93
