Tracy Narine | 10403ec | 2023-11-28 11:55:08 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2024 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include <armnn/Exceptions.hpp> |
| 7 | |
| 8 | #pragma once |
| 9 | |
| 10 | inline void CreateRescaleTosaOperator(const std::string& inputName, |
| 11 | const std::string& outputName, |
| 12 | DType output_type, |
| 13 | const std::vector<int32_t>& shape, |
John Mcloughlin | ceb4428 | 2024-04-23 16:47:04 +0100 | [diff] [blame^] | 14 | const std::vector<int32_t>& multipliers, |
| 15 | const std::vector<int32_t>& shifts, |
Tracy Narine | 10403ec | 2023-11-28 11:55:08 +0000 | [diff] [blame] | 16 | int32_t input_zp, |
| 17 | int32_t output_zp, |
| 18 | bool double_round, |
| 19 | bool scale32, |
John Mcloughlin | ceb4428 | 2024-04-23 16:47:04 +0100 | [diff] [blame^] | 20 | bool per_channel, |
Tracy Narine | 10403ec | 2023-11-28 11:55:08 +0000 | [diff] [blame] | 21 | TosaSerializationOperator** op, |
| 22 | TosaSerializationTensor** tensor) |
| 23 | { |
| 24 | if (!op) |
| 25 | { |
| 26 | throw armnn::Exception("CreateRescaleTosaOperator: nullptr op"); |
| 27 | } |
| 28 | |
Tracy Narine | 10403ec | 2023-11-28 11:55:08 +0000 | [diff] [blame] | 29 | TosaRescaleAttribute attribute(input_zp, |
| 30 | output_zp, |
| 31 | multipliers, |
| 32 | shifts, |
| 33 | scale32, |
| 34 | double_round, |
John Mcloughlin | ceb4428 | 2024-04-23 16:47:04 +0100 | [diff] [blame^] | 35 | per_channel, |
Teresa Charlin | 571a4f7 | 2024-03-26 11:18:42 +0000 | [diff] [blame] | 36 | false, // input_unsigned |
| 37 | false); // output_unsigned |
Tracy Narine | 10403ec | 2023-11-28 11:55:08 +0000 | [diff] [blame] | 38 | |
| 39 | // op |
| 40 | *op = new TosaSerializationOperator(Op_RESCALE, Attribute_RescaleAttribute, &attribute, {inputName}, {outputName}); |
| 41 | if (!(*op)) |
| 42 | { |
| 43 | throw armnn::Exception("CreateRescaleTosaOperator: failed to created operator"); |
| 44 | } |
| 45 | if (tensor != nullptr) |
| 46 | { |
| 47 | // tensor |
| 48 | *tensor = new TosaSerializationTensor(outputName, shape, output_type, {}); |
| 49 | if (! (*tensor)) |
| 50 | { |
| 51 | throw armnn::Exception("CreateRescaleTosaOperator: failed to created tensor"); |
| 52 | } |
| 53 | } |
| 54 | } |
| 55 | |
| 56 | inline void CreateRescaleTosaOperator(const std::string& inputName, |
| 57 | const std::string& outputName, |
| 58 | DType output_type, |
| 59 | const std::vector<int32_t>& shape, |
John Mcloughlin | ceb4428 | 2024-04-23 16:47:04 +0100 | [diff] [blame^] | 60 | int32_t scale_multiplier, |
| 61 | int32_t scale_shift, |
| 62 | int32_t input_zp, |
| 63 | int32_t output_zp, |
| 64 | bool double_round, |
| 65 | bool scale32, |
| 66 | bool per_channel, |
| 67 | TosaSerializationOperator** op, |
| 68 | TosaSerializationTensor** tensor) |
| 69 | { |
| 70 | const std::vector<int32_t> multipliers{scale_multiplier}; |
| 71 | const std::vector<int32_t> shifts{scale_shift}; |
| 72 | CreateRescaleTosaOperator(inputName, outputName, output_type, shape, multipliers, shifts, |
| 73 | input_zp, output_zp, double_round, scale32, per_channel, op, tensor); |
| 74 | } |
| 75 | |
| 76 | /// The following is taken from mlir/lib/Dialect/Tosa/Utils/QuantUtils.cpp in the LLVM project |
| 77 | /// From a scale value, generates multiplier and shift values where |
| 78 | /// mantissa is in [-1.0,-0.5] or [0.5, 1.0] such that |
| 79 | /// multiplier = mantissa*2^shift for 32-bit scaling. |
| 80 | static void ComputeMultiplierAndShiftTosaScale32(double scale, |
| 81 | int32_t &multiplier, |
| 82 | int32_t &shift) |
| 83 | { |
| 84 | const double mantissa = std::frexp(scale, &shift); |
| 85 | auto shiftedM = std::round(mantissa * (int64_t(1) << 31)); |
| 86 | |
| 87 | // Can't be greater than 1.0. |
| 88 | if (!(shiftedM <= (int64_t(1) << 31))) |
| 89 | { |
| 90 | throw armnn::Exception("Shifted mantissa exceeds 32 signed bits"); |
| 91 | } |
| 92 | |
| 93 | if (shiftedM == (int64_t(1) << 31)) |
| 94 | { |
| 95 | shiftedM /= 2; |
| 96 | shift++; |
| 97 | } |
| 98 | |
| 99 | // TOSA expects right shift to be positive, and embed (1 << 31) into right |
| 100 | // shift bits. |
| 101 | shift = (-shift) + 31; |
| 102 | |
| 103 | if (!(shiftedM <= std::numeric_limits<int32_t>::max())) |
| 104 | { |
| 105 | throw armnn::Exception("Shifted mantissa exceeds 32-bit signed output type"); |
| 106 | } |
| 107 | |
| 108 | multiplier = static_cast<int32_t>(shiftedM); |
| 109 | |
| 110 | // Shifting tops out at 62 bits. Right shift to make 62 bits the max. |
| 111 | // The limit of 62 on shift allows the shift to be decomposed as |
| 112 | // two right shifts of 31. |
| 113 | if (shift > 62) |
| 114 | { |
| 115 | // Shifting the multiplier by more than 32-bits is unnecessary. |
| 116 | multiplier = multiplier >> std::min<int32_t>(31, shift - 62); |
| 117 | shift = 62; |
| 118 | } |
| 119 | } |
| 120 | |
| 121 | /// The following is taken from mlir/lib/Dialect/Tosa/Utils/QuantUtils.cpp in the LLVM project |
| 122 | /// From a scale value, generates multiplier and shift values where |
| 123 | /// mantissa is in [-1.0,-0.5] or [0.5, 1.0] such that |
| 124 | /// multiplier = mantissa*2^shift for 16-bit scaling. |
| 125 | static void ComputeMultiplierAndShiftTosaScale16(double scale, |
| 126 | int32_t &multiplier, |
| 127 | int32_t &shift) |
| 128 | { |
| 129 | const double mantissa = std::frexp(scale, &shift); |
| 130 | auto shiftedM = std::round(mantissa * (int64_t(1) << 15)); |
| 131 | |
| 132 | // Can't be greater than 1.0. |
| 133 | if (!(shiftedM <= (int64_t(1) << 15))) |
| 134 | { |
| 135 | throw armnn::Exception("Shifted mantissa exceeds 16 signed bits"); |
| 136 | } |
| 137 | |
| 138 | if (shiftedM == (int64_t(1) << 15)) |
| 139 | { |
| 140 | shiftedM /= 2; |
| 141 | shift++; |
| 142 | } |
| 143 | |
| 144 | // TOSA expects right shift to be positive and embed (1 << 15) into right |
| 145 | // shift bits. |
| 146 | shift = (-shift) + 15; |
| 147 | |
| 148 | if (!(shiftedM <= std::numeric_limits<int32_t>::max())) |
| 149 | { |
| 150 | throw armnn::Exception("Shifted mantissa exceeds 32-bit signed output type"); |
| 151 | } |
| 152 | |
| 153 | multiplier = static_cast<int32_t>(shiftedM); |
| 154 | |
| 155 | // Shifting tops out at 62 bits. Right shift to make 62 bits the max. |
| 156 | // The limit of 62 on shift allows the shift to be decomposed as |
| 157 | // two right shifts of 31. |
| 158 | if (shift > 62) |
| 159 | { |
| 160 | // Shifting the multiplier by more than 31-bits is unnecessary. |
| 161 | multiplier = multiplier >> std::min<int32_t>(31, shift - 62); |
| 162 | shift = 62; |
| 163 | } |
| 164 | } |
| 165 | |
| 166 | inline void CreateRescaleTosaOperator(const std::string& inputName, |
| 167 | const std::string& outputName, |
| 168 | DType output_type, |
| 169 | const std::vector<int32_t>& shape, |
Tracy Narine | 10403ec | 2023-11-28 11:55:08 +0000 | [diff] [blame] | 170 | double scale, |
| 171 | int32_t input_zp, |
| 172 | int32_t output_zp, |
| 173 | bool double_round, |
| 174 | bool scale32, |
| 175 | TosaSerializationOperator** op, |
| 176 | TosaSerializationTensor** tensor) |
| 177 | { |
Tracy Narine | 10403ec | 2023-11-28 11:55:08 +0000 | [diff] [blame] | 178 | int32_t multiplier; |
| 179 | int32_t shift; |
John Mcloughlin | ceb4428 | 2024-04-23 16:47:04 +0100 | [diff] [blame^] | 180 | |
| 181 | if (scale32) |
| 182 | { |
| 183 | ComputeMultiplierAndShiftTosaScale32(scale, multiplier, shift); |
| 184 | } |
| 185 | else |
| 186 | { |
| 187 | ComputeMultiplierAndShiftTosaScale16(scale, multiplier, shift); |
| 188 | } |
| 189 | |
Tracy Narine | 10403ec | 2023-11-28 11:55:08 +0000 | [diff] [blame] | 190 | CreateRescaleTosaOperator(inputName, outputName, output_type, shape, multiplier, shift, |
John Mcloughlin | ceb4428 | 2024-04-23 16:47:04 +0100 | [diff] [blame^] | 191 | input_zp, output_zp, double_round, scale32, false, op, tensor); |
| 192 | } |
| 193 | |
| 194 | inline void CreateRescaleTosaOperatorPerChannel(const std::string& inputName, |
| 195 | const std::string& outputName, |
| 196 | DType output_type, |
| 197 | const std::vector<int32_t>& shape, |
| 198 | int32_t input_zp, |
| 199 | int32_t output_zp, |
| 200 | bool double_round, |
| 201 | bool scale32, |
| 202 | double input_scale, |
| 203 | double output_scale, |
| 204 | const std::vector<float>& weight_scales, |
| 205 | TosaSerializationOperator** op, |
| 206 | TosaSerializationTensor** tensor) |
| 207 | { |
| 208 | std::vector<int32_t> op_tensor_multipliers; |
| 209 | std::vector<int32_t> op_tensor_shifts; |
| 210 | op_tensor_multipliers.reserve(weight_scales.size()); |
| 211 | op_tensor_shifts.reserve(weight_scales.size()); |
| 212 | |
| 213 | for (const float& weight_scale : weight_scales) |
| 214 | { |
| 215 | double op_tensor_scale = (input_scale * weight_scale) / output_scale; |
| 216 | int32_t multiplier; |
| 217 | int32_t shift; |
| 218 | |
| 219 | if (scale32) |
| 220 | { |
| 221 | ComputeMultiplierAndShiftTosaScale32(op_tensor_scale, multiplier, shift); |
| 222 | } |
| 223 | else |
| 224 | { |
| 225 | ComputeMultiplierAndShiftTosaScale16(op_tensor_scale, multiplier, shift); |
| 226 | } |
| 227 | |
| 228 | op_tensor_multipliers.push_back(multiplier); |
| 229 | op_tensor_shifts.push_back(shift); |
| 230 | } |
| 231 | |
| 232 | CreateRescaleTosaOperator(inputName, outputName, output_type, shape, op_tensor_multipliers, op_tensor_shifts, |
| 233 | input_zp, output_zp, double_round, scale32, true, op, tensor); |
Tracy Narine | 10403ec | 2023-11-28 11:55:08 +0000 | [diff] [blame] | 234 | } |
| 235 | |
| 236 | inline void CreateFromInt32RescaleTosaOperator(const std::string& inputName, |
| 237 | const std::string& outputName, |
John Mcloughlin | ceb4428 | 2024-04-23 16:47:04 +0100 | [diff] [blame^] | 238 | DType output_type, |
| 239 | const std::vector<int32_t>& shape, |
| 240 | double output_scale, |
| 241 | int32_t output_zp, |
| 242 | TosaSerializationOperator** op, |
| 243 | TosaSerializationTensor** tensor) |
Tracy Narine | 10403ec | 2023-11-28 11:55:08 +0000 | [diff] [blame] | 244 | { |
John Mcloughlin | ceb4428 | 2024-04-23 16:47:04 +0100 | [diff] [blame^] | 245 | CreateRescaleTosaOperator(inputName, outputName, output_type, shape, output_scale, |
| 246 | 0, output_zp, true, true, op, tensor); |
Tracy Narine | 10403ec | 2023-11-28 11:55:08 +0000 | [diff] [blame] | 247 | } |