| // |
| // Copyright © 2017 Arm Ltd. All rights reserved. |
| // SPDX-License-Identifier: MIT |
| // |
| |
| #include "ConvImpl.hpp" |
| |
| #include <armnn/utility/Assert.hpp> |
| |
| #include <cmath> |
| #include <limits> |
| |
| namespace armnn |
| { |
| |
| QuantizedMultiplierSmallerThanOne::QuantizedMultiplierSmallerThanOne(float multiplier) |
| { |
| ARMNN_ASSERT(multiplier >= 0.0f && multiplier < 1.0f); |
| if (multiplier == 0.0f) |
| { |
| m_Multiplier = 0; |
| m_RightShift = 0; |
| } |
| else |
| { |
| const double q = std::frexp(multiplier, &m_RightShift); |
| m_RightShift = -m_RightShift; |
| int64_t qFixed = static_cast<int64_t>(std::round(q * (1ll << 31))); |
| ARMNN_ASSERT(qFixed <= (1ll << 31)); |
| if (qFixed == (1ll << 31)) |
| { |
| qFixed /= 2; |
| --m_RightShift; |
| } |
| ARMNN_ASSERT(m_RightShift >= 0); |
| ARMNN_ASSERT(qFixed <= std::numeric_limits<int32_t>::max()); |
| m_Multiplier = static_cast<int32_t>(qFixed); |
| } |
| } |
| |
| int32_t QuantizedMultiplierSmallerThanOne::operator*(int32_t rhs) const |
| { |
| int32_t x = SaturatingRoundingDoublingHighMul(rhs, m_Multiplier); |
| return RoundingDivideByPOT(x, m_RightShift); |
| } |
| |
| int32_t QuantizedMultiplierSmallerThanOne::SaturatingRoundingDoublingHighMul(int32_t a, int32_t b) |
| { |
| // Check for overflow. |
| if (a == b && a == std::numeric_limits<int32_t>::min()) |
| { |
| return std::numeric_limits<int32_t>::max(); |
| } |
| int64_t a_64(a); |
| int64_t b_64(b); |
| int64_t ab_64 = a_64 * b_64; |
| int32_t nudge = ab_64 >= 0 ? (1 << 30) : (1 - (1 << 30)); |
| int32_t ab_x2_high32 = static_cast<std::int32_t>((ab_64 + nudge) / (1ll << 31)); |
| return ab_x2_high32; |
| } |
| |
| int32_t QuantizedMultiplierSmallerThanOne::RoundingDivideByPOT(int32_t x, int exponent) |
| { |
| ARMNN_ASSERT(exponent >= 0 && exponent <= 31); |
| int32_t mask = (1 << exponent) - 1; |
| int32_t remainder = x & mask; |
| int32_t threshold = (mask >> 1) + (x < 0 ? 1 : 0); |
| return (x >> exponent) + (remainder > threshold ? 1 : 0); |
| } |
| |
| void Convolve(const TensorShape& rInputShape, |
| Decoder<float>& rInputDecoder, |
| const TensorShape& rOutputShape, |
| Encoder<float>& rOutputEncoder, |
| const TensorShape& rFilterShape, |
| Decoder<float>& rFilterDecoder, |
| bool biasEnabled, |
| Decoder<float>* pBiasDecoder, |
| DataLayout dataLayout, |
| unsigned int paddingTop, |
| unsigned int paddingLeft, |
| unsigned int xStride, |
| unsigned int yStride, |
| unsigned int xDilation, |
| unsigned int yDilation, |
| bool depthwise) |
| { |
| if (biasEnabled && !pBiasDecoder) |
| { |
| throw InvalidArgumentException("Bias is enabled but the bias data is invalid"); |
| } |
| const armnnUtils::DataLayoutIndexed dataLayoutIndexed(dataLayout); |
| |
| const unsigned int channelsIndex = dataLayoutIndexed.GetChannelsIndex(); |
| const unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| const unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| |
| const unsigned int depthMultiplier = depthwise ? rFilterShape[0] : 1; |
| const unsigned int inputChannels = depthwise ? rFilterShape[1] : rFilterShape[channelsIndex]; |
| const unsigned int outputChannels = depthwise ? inputChannels * depthMultiplier : rFilterShape[0]; |
| |
| const unsigned int batchSize = rOutputShape[0]; |
| const unsigned int outputHeight = rOutputShape[heightIndex]; |
| const unsigned int outputWidth = rOutputShape[widthIndex]; |
| const unsigned int inputHeight = rInputShape[heightIndex]; |
| const unsigned int inputWidth = rInputShape[widthIndex]; |
| |
| const unsigned int filterHeight = depthwise ? rFilterShape[2] : rFilterShape[heightIndex]; |
| const unsigned int filterWidth = depthwise ? rFilterShape[3] : rFilterShape[widthIndex]; |
| |
| const std::vector<float> inputVec = rInputDecoder.DecodeTensor(rInputShape); |
| const std::vector<float> filterVec = rFilterDecoder.DecodeTensor(rFilterShape, depthMultiplier, depthwise); |
| |
| const TensorShape biasShape{outputChannels}; |
| const std::vector<float> biasVec = biasEnabled ? pBiasDecoder->DecodeTensor(biasShape) : std::vector<float>(); |
| |
| unsigned int depthwiseMultiplierIdx = 0; |
| for (unsigned int batchIdx = 0; batchIdx < batchSize; batchIdx++) |
| { |
| for (unsigned int cOutput = 0; cOutput < outputChannels; cOutput++) |
| { |
| for (unsigned int yOutput = 0; yOutput < outputHeight; yOutput++) |
| { |
| for (unsigned int xOutput = 0; xOutput < outputWidth; xOutput++) |
| { |
| // This loop goes over each output element. |
| float sum = 0.0f; |
| |
| // For depthwise, each output channel corresponds to exactly one input channel. |
| // For normal, must loop over each input channel. |
| for (unsigned int cInput = 0; cInput < (depthwise ? 1 : inputChannels); cInput++) |
| { |
| if (depthwise) |
| { |
| depthwiseMultiplierIdx = 0; |
| cInput = cOutput / depthMultiplier; |
| depthwiseMultiplierIdx = cOutput % depthMultiplier; |
| } |
| |
| for (unsigned int yFilter = 0; yFilter < filterHeight; yFilter++) |
| { |
| for (unsigned int xFilter = 0; xFilter < filterWidth; xFilter++) |
| { |
| // This loop goes over each input element for each output element. |
| unsigned int filterIndex = 0; |
| |
| // Since dimensionality of kernel depends on depthwiseness, so does index. |
| if (depthwise) |
| { |
| filterIndex = depthwiseMultiplierIdx * filterWidth * filterHeight * inputChannels + |
| cInput * filterWidth * filterHeight + |
| yFilter * filterWidth + |
| xFilter; |
| } |
| else |
| { |
| // Keep this implementation, as using DataLayoutIndexed::GetIndex causes great |
| // performance regression. |
| if (dataLayoutIndexed.GetDataLayout() == DataLayout::NHWC) |
| { |
| filterIndex = cOutput * filterHeight * filterWidth * inputChannels + |
| yFilter * filterWidth * inputChannels + |
| xFilter * inputChannels + |
| cInput; |
| } |
| else |
| { |
| filterIndex = cOutput * filterWidth * filterHeight * inputChannels + |
| cInput * filterWidth * filterHeight + |
| yFilter * filterWidth + |
| xFilter; |
| } |
| } |
| |
| unsigned int yInput = yOutput * yStride + yFilter * yDilation; |
| unsigned int xInput = xOutput * xStride + xFilter * xDilation; |
| |
| float inputValue; |
| |
| // Check if we're in the padding. |
| if (yInput < paddingTop || yInput >= inputHeight + paddingTop || |
| xInput < paddingLeft || xInput >= inputWidth + paddingLeft) |
| { |
| inputValue = 0.0f; |
| } |
| else |
| { |
| unsigned int inputIndex = 0; |
| |
| // Keep this implementation, as using DataLayoutIndexed::GetIndex causes great |
| // performance regression. |
| if (dataLayoutIndexed.GetDataLayout() == DataLayout::NHWC) |
| { |
| inputIndex = batchIdx * inputHeight * inputWidth * inputChannels + |
| (yInput - paddingTop) * inputWidth * inputChannels + |
| (xInput - paddingLeft) * inputChannels + |
| cInput; |
| } |
| else |
| { |
| inputIndex = batchIdx * inputWidth * inputHeight * inputChannels + |
| inputWidth * inputHeight * cInput + |
| inputWidth * (yInput - paddingTop) + |
| xInput - paddingLeft; |
| } |
| inputValue = inputVec[inputIndex]; |
| } |
| |
| sum += filterVec[filterIndex] * inputValue; |
| } |
| } |
| } |
| |
| if (biasEnabled) |
| { |
| sum += biasVec[cOutput]; |
| } |
| |
| unsigned int outIdx; |
| if (dataLayoutIndexed.GetDataLayout() == DataLayout::NHWC) |
| { |
| outIdx = batchIdx * outputHeight * outputWidth * outputChannels + |
| yOutput * outputWidth * outputChannels + |
| xOutput * outputChannels + |
| cOutput; |
| } |
| else |
| { |
| outIdx = batchIdx * outputHeight * outputWidth * outputChannels + |
| cOutput * outputHeight * outputWidth + |
| yOutput * outputWidth + |
| xOutput; |
| } |
| |
| rOutputEncoder[outIdx]; |
| rOutputEncoder.Set(sum); |
| } |
| } |
| } |
| } |
| } |
| |
| } // namespace armnn |