| 16 | { |
| 17 | |
| 18 | void BatchNormImpl(const BatchNormalizationQueueDescriptor& data, |
| 19 | Decoder<float>& meanDecoder, |
| 20 | Decoder<float>& varianceDecoder, |
| 21 | Decoder<float>& betaDecoder, |
| 22 | Decoder<float>& gammaDecoder, |
| 23 | Decoder<float>& inputDecoder, |
| 24 | Encoder<float>& outputEncoder) |
| 25 | { |
| 26 | const TensorInfo& inputInfo = GetTensorInfo(data.m_Inputs[0]); |
| 27 | const TensorShape inputShape = inputInfo.GetShape(); |
| 28 | |
| 29 | armnnUtils::DataLayoutIndexed dataLayout(data.m_Parameters.m_DataLayout); |
| 30 | |
| 31 | unsigned int inputBatches = inputShape[0]; |
| 32 | unsigned int inputHeight = inputShape[dataLayout.GetHeightIndex()]; |
| 33 | unsigned int inputWidth = inputShape[dataLayout.GetWidthIndex()]; |
| 34 | unsigned int inputChannels = inputShape[dataLayout.GetChannelsIndex()]; |
| 35 | |
| 36 | for (unsigned int c = 0; c < inputChannels; c++) |
| 37 | { |
| 38 | meanDecoder[c]; |
| 39 | varianceDecoder[c]; |
| 40 | betaDecoder[c]; |
| 41 | gammaDecoder[c]; |
| 42 | float mean = meanDecoder.Get(); |
| 43 | float var = varianceDecoder.Get(); |
| 44 | float beta = betaDecoder.Get(); |
| 45 | float gamma = gammaDecoder.Get(); |
| 46 | |
| 47 | float mult = gamma / sqrtf(var + data.m_Parameters.m_Eps); |
| 48 | float add = beta - mult * mean; |
| 49 | |
| 50 | for (unsigned int n = 0; n < inputBatches; n++) |
| 51 | { |
| 52 | for (unsigned int h = 0; h < inputHeight; h++) |
| 53 | { |
| 54 | for (unsigned int w = 0; w < inputWidth; w++) |
| 55 | { |
| 56 | unsigned int index = dataLayout.GetIndex(inputShape, n, c, h, w); |
| 57 | inputDecoder[index]; |
| 58 | outputEncoder[index]; |
| 59 | outputEncoder.Set(mult * inputDecoder.Get() + add); |
| 60 | } |
| 61 | } |
| 62 | } |
| 63 | } |
| 64 | } |
| 65 | |
| 66 | } // namespace armnn |
no test coverage detected