Samuel Yap | 6b47809 | 2022-07-06 15:36:03 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2022 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "BatchMatMulImpl.hpp" |
| 7 | |
| 8 | #include <armnn/backends/WorkloadData.hpp> |
| 9 | #include <armnn/Logging.hpp> |
| 10 | |
| 11 | namespace armnn |
| 12 | { |
| 13 | |
| 14 | void BatchMatMul::BatchMatMulImpl() |
| 15 | { |
| 16 | inputXData = inputXDecoder.DecodeTensor(inputXInfo.GetShape()); |
| 17 | inputYData = inputYDecoder.DecodeTensor(inputYInfo.GetShape()); |
| 18 | // At this point, we don't touch the input decoders - just the resultant vectors |
| 19 | |
| 20 | // Pre-transpose and pre-adjoint if their vectors aren't empty |
| 21 | // and also DataLayouts which may change with permutations/adjoints |
| 22 | |
| 23 | // Todo: Have you updated input validation and inferred output shapes to accommodate for these pre-permutes? |
| 24 | |
| 25 | auto idx = std::vector<unsigned int>(outputInfo.GetNumDimensions(), 0); |
| 26 | RecurseBMM(idx, 0); |
| 27 | } |
| 28 | |
| 29 | void BatchMatMul::RecurseBMM(std::vector<unsigned int>& curIdx, unsigned int curDim) |
| 30 | { |
| 31 | // We're working off of the indexes of the output tensor (the max possible shape) |
| 32 | |
| 33 | if(!(curDim < outputInfo.GetNumDimensions())) |
| 34 | { |
| 35 | // We're at the leaf level of this call tree, so we operate here (each leaf is a data point) |
| 36 | |
| 37 | auto axesToMul = BatchMatMulDescriptor::GetAxesToMul(params, |
| 38 | inputXInfo.GetShape(), |
| 39 | inputYInfo.GetShape()); |
| 40 | AdjustAxesToMulForUnequalRanks(axesToMul); |
| 41 | |
| 42 | unsigned int inputXColDim = axesToMul.first.second; |
| 43 | unsigned int inputYRowDim = axesToMul.second.first; |
| 44 | |
| 45 | unsigned int inputYRowSize = inputYInfo.GetShape()[inputYRowDim]; |
| 46 | |
| 47 | float sum = 0.0f; |
| 48 | |
| 49 | // You could also use inputXColSize |
| 50 | for (unsigned int inputYRowIdx = 0; inputYRowIdx < inputYRowSize; inputYRowIdx++) { |
| 51 | auto xIdx = curIdx; |
| 52 | xIdx[inputXColDim] = inputYRowIdx; |
| 53 | |
| 54 | auto yIdx = curIdx; |
| 55 | yIdx[inputYRowDim] = inputYRowIdx; |
| 56 | |
| 57 | sum += (GetValueAt(DataSlot::InputX, xIdx) |
| 58 | * GetValueAt(DataSlot::InputY, yIdx)); |
| 59 | } |
| 60 | |
| 61 | SetValueAt(sum, DataSlot::Output, curIdx); |
| 62 | |
| 63 | return; |
| 64 | } |
| 65 | |
| 66 | for (unsigned int i = 0; i < outputInfo.GetShape()[curDim]; i++) |
| 67 | { |
| 68 | curIdx[curDim] = i; |
| 69 | RecurseBMM(curIdx, curDim+1); |
| 70 | } |
| 71 | } |
| 72 | |
| 73 | void BatchMatMul::AdjustAxesToMulForUnequalRanks( |
| 74 | std::pair<std::pair<unsigned int, unsigned int>, std::pair<unsigned int, unsigned int>>& axesToMul) |
| 75 | { |
| 76 | long rankDiff = static_cast<long>(inputXInfo.GetNumDimensions()) - inputYInfo.GetNumDimensions(); |
| 77 | if(rankDiff == 0) |
| 78 | { |
| 79 | return; |
| 80 | } |
| 81 | else if(rankDiff < 0) |
| 82 | { |
| 83 | // Y is the larger one |
| 84 | axesToMul.first.first += static_cast<std::make_unsigned<unsigned int>::type>(std::abs(rankDiff)); |
| 85 | axesToMul.first.second += static_cast<std::make_unsigned<unsigned int>::type>(std::abs(rankDiff)); |
| 86 | } |
| 87 | else if(rankDiff > 0) |
| 88 | { |
| 89 | // X is the larger one |
| 90 | axesToMul.second.first += static_cast<std::make_unsigned<unsigned int>::type>(std::abs(rankDiff)); |
| 91 | axesToMul.second.second += static_cast<std::make_unsigned<unsigned int>::type>(std::abs(rankDiff)); |
| 92 | } |
| 93 | } |
| 94 | |
| 95 | float BatchMatMul::GetValueAt(DataSlot type, std::vector<unsigned int> idx) |
| 96 | { |
| 97 | // This gets the data from the input vector that we have, Not the decoder |
| 98 | // But for the output, it is operating on the encoder itself |
| 99 | |
| 100 | AdjustToSafeIdx(type, idx); |
| 101 | unsigned int flatIdx = CalcFlatIdx(type, idx); |
| 102 | float value = 0.0f; |
| 103 | |
| 104 | switch(type) |
| 105 | { |
| 106 | case DataSlot::InputX: |
| 107 | value = inputXData[flatIdx]; |
| 108 | break; |
| 109 | case DataSlot::InputY: |
| 110 | value = inputYData[flatIdx]; |
| 111 | break; |
| 112 | case DataSlot::Output: |
| 113 | outputEncoder[flatIdx]; |
| 114 | value = outputEncoder.Get(); |
| 115 | break; |
| 116 | default: |
| 117 | break; |
| 118 | } |
| 119 | |
| 120 | return value; |
| 121 | } |
| 122 | |
| 123 | void BatchMatMul::SetValueAt(float value, DataSlot type, std::vector<unsigned int> idx) |
| 124 | { |
| 125 | AdjustToSafeIdx(type, idx); |
| 126 | |
| 127 | unsigned int flatIdx = CalcFlatIdx(type, idx); |
| 128 | |
| 129 | switch(type) |
| 130 | { |
| 131 | case DataSlot::InputX: |
| 132 | inputXData[flatIdx] = value; |
| 133 | break; |
| 134 | case DataSlot::InputY: |
| 135 | inputYData[flatIdx] = value; |
| 136 | break; |
| 137 | case DataSlot::Output: |
| 138 | outputEncoder[flatIdx]; |
| 139 | outputEncoder.Set(value); |
| 140 | break; |
| 141 | default: |
| 142 | break; |
| 143 | } |
| 144 | } |
| 145 | |
| 146 | void BatchMatMul::AdjustToSafeIdx(DataSlot type, std::vector<unsigned int>& idx) |
| 147 | { |
| 148 | for(unsigned int dim = 0; dim < idx.size(); dim++) |
| 149 | { |
| 150 | switch(type) |
| 151 | { |
| 152 | case DataSlot::InputX: |
| 153 | { |
| 154 | auto xRank = inputXInfo.GetNumDimensions(); |
| 155 | auto xDiff = outputInfo.GetNumDimensions() - xRank; |
| 156 | if (dim < xDiff || |
| 157 | idx[dim] > inputXInfo.GetShape()[dim-xDiff]-1) |
| 158 | { |
| 159 | idx[dim] = 0; // Broadcasting |
| 160 | } |
| 161 | break; |
| 162 | } |
| 163 | case DataSlot::InputY: |
| 164 | { |
| 165 | auto yRank = inputYInfo.GetNumDimensions(); |
| 166 | auto yDiff = outputInfo.GetNumDimensions() - yRank; |
| 167 | if (dim < yDiff || |
| 168 | idx[dim] > inputYInfo.GetShape()[dim-yDiff]-1) |
| 169 | { |
| 170 | idx[dim] = 0; |
| 171 | } |
| 172 | break; |
| 173 | } |
| 174 | case DataSlot::Output: |
| 175 | { |
| 176 | // Our indices are based off the output |
| 177 | break; |
| 178 | } |
| 179 | default: |
| 180 | break; |
| 181 | } |
| 182 | } |
| 183 | } |
| 184 | |
| 185 | unsigned int BatchMatMul::CalcFlatIdx(DataSlot type, const std::vector<unsigned int>& idx) |
| 186 | { |
| 187 | unsigned int result = idx[idx.size()-1]; |
| 188 | |
| 189 | unsigned int dimMultiplier = 1; |
| 190 | |
| 191 | unsigned int offset; |
| 192 | |
| 193 | // -2 because final dim is already accounted for in the multiplier (last dim is just a multiplier of 1x) |
| 194 | for(unsigned int i = static_cast<unsigned int>(idx.size()-2); static_cast<int>(i) >= 0; i--) |
| 195 | { |
| 196 | switch(type) |
| 197 | { |
| 198 | case DataSlot::InputX: |
| 199 | offset = outputInfo.GetNumDimensions() - inputXInfo.GetNumDimensions(); |
| 200 | dimMultiplier *= inputXInfo.GetShape()[i + 1 - offset]; |
| 201 | break; |
| 202 | case DataSlot::InputY: |
| 203 | offset = outputInfo.GetNumDimensions() - inputYInfo.GetNumDimensions(); |
| 204 | dimMultiplier *= inputYInfo.GetShape()[i + 1 - offset]; |
| 205 | break; |
| 206 | case DataSlot::Output: |
| 207 | dimMultiplier *= outputInfo.GetShape()[i+1]; |
| 208 | break; |
| 209 | default: |
| 210 | break; |
| 211 | } |
| 212 | result += (idx[i] * dimMultiplier); |
| 213 | } |
| 214 | return result; |
| 215 | } |
| 216 | |
| 217 | template <typename T> |
| 218 | std::string BatchMatMul::StringifyVec(const std::vector<T>& vec) |
| 219 | { |
| 220 | std::string res = "{ "; |
| 221 | for(auto x : vec) |
| 222 | { |
| 223 | res += std::to_string(x); |
| 224 | res += " "; |
| 225 | } |
| 226 | res += "}"; |
| 227 | return res; |
| 228 | } |
| 229 | |
| 230 | } // namespace armnn |