Laurent Carlier | 749294b | 2020-06-01 09:03:17 +0100 | [diff] [blame] | 1 | // |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
David Beck | ecb56cd | 2018-09-05 12:52:57 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4 | // |
| 5 | |
| 6 | #include "ConvImpl.hpp" |
| 7 | |
Narumol Prangnawarat | ac2770a | 2020-04-01 16:51:23 +0100 | [diff] [blame] | 8 | #include <armnn/utility/Assert.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 9 | |
| 10 | #include <cmath> |
| 11 | #include <limits> |
| 12 | |
| 13 | namespace armnn |
| 14 | { |
| 15 | |
| 16 | QuantizedMultiplierSmallerThanOne::QuantizedMultiplierSmallerThanOne(float multiplier) |
| 17 | { |
Narumol Prangnawarat | ac2770a | 2020-04-01 16:51:23 +0100 | [diff] [blame] | 18 | ARMNN_ASSERT(multiplier >= 0.0f && multiplier < 1.0f); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 19 | if (multiplier == 0.0f) |
| 20 | { |
| 21 | m_Multiplier = 0; |
| 22 | m_RightShift = 0; |
| 23 | } |
| 24 | else |
| 25 | { |
| 26 | const double q = std::frexp(multiplier, &m_RightShift); |
| 27 | m_RightShift = -m_RightShift; |
| 28 | int64_t qFixed = static_cast<int64_t>(std::round(q * (1ll << 31))); |
Narumol Prangnawarat | ac2770a | 2020-04-01 16:51:23 +0100 | [diff] [blame] | 29 | ARMNN_ASSERT(qFixed <= (1ll << 31)); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 30 | if (qFixed == (1ll << 31)) |
| 31 | { |
| 32 | qFixed /= 2; |
| 33 | --m_RightShift; |
| 34 | } |
Narumol Prangnawarat | ac2770a | 2020-04-01 16:51:23 +0100 | [diff] [blame] | 35 | ARMNN_ASSERT(m_RightShift >= 0); |
| 36 | ARMNN_ASSERT(qFixed <= std::numeric_limits<int32_t>::max()); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 37 | m_Multiplier = static_cast<int32_t>(qFixed); |
| 38 | } |
| 39 | } |
| 40 | |
| 41 | int32_t QuantizedMultiplierSmallerThanOne::operator*(int32_t rhs) const |
| 42 | { |
| 43 | int32_t x = SaturatingRoundingDoublingHighMul(rhs, m_Multiplier); |
| 44 | return RoundingDivideByPOT(x, m_RightShift); |
| 45 | } |
| 46 | |
| 47 | int32_t QuantizedMultiplierSmallerThanOne::SaturatingRoundingDoublingHighMul(int32_t a, int32_t b) |
| 48 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 49 | // Check for overflow. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 50 | if (a == b && a == std::numeric_limits<int32_t>::min()) |
| 51 | { |
| 52 | return std::numeric_limits<int32_t>::max(); |
| 53 | } |
| 54 | int64_t a_64(a); |
| 55 | int64_t b_64(b); |
| 56 | int64_t ab_64 = a_64 * b_64; |
| 57 | int32_t nudge = ab_64 >= 0 ? (1 << 30) : (1 - (1 << 30)); |
| 58 | int32_t ab_x2_high32 = static_cast<std::int32_t>((ab_64 + nudge) / (1ll << 31)); |
| 59 | return ab_x2_high32; |
| 60 | } |
| 61 | |
| 62 | int32_t QuantizedMultiplierSmallerThanOne::RoundingDivideByPOT(int32_t x, int exponent) |
| 63 | { |
Narumol Prangnawarat | ac2770a | 2020-04-01 16:51:23 +0100 | [diff] [blame] | 64 | ARMNN_ASSERT(exponent >= 0 && exponent <= 31); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 65 | int32_t mask = (1 << exponent) - 1; |
| 66 | int32_t remainder = x & mask; |
| 67 | int32_t threshold = (mask >> 1) + (x < 0 ? 1 : 0); |
| 68 | return (x >> exponent) + (remainder > threshold ? 1 : 0); |
| 69 | } |
| 70 | |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 71 | void Convolve(const TensorShape& rInputShape, |
| 72 | Decoder<float>& rInputDecoder, |
| 73 | const TensorShape& rOutputShape, |
| 74 | Encoder<float>& rOutputEncoder, |
| 75 | const TensorShape& rFilterShape, |
| 76 | Decoder<float>& rFilterDecoder, |
| 77 | bool biasEnabled, |
| 78 | Decoder<float>* pBiasDecoder, |
| 79 | DataLayout dataLayout, |
| 80 | unsigned int paddingTop, |
| 81 | unsigned int paddingLeft, |
| 82 | unsigned int xStride, |
| 83 | unsigned int yStride, |
| 84 | unsigned int xDilation, |
| 85 | unsigned int yDilation, |
| 86 | bool depthwise) |
| 87 | { |
| 88 | if (biasEnabled && !pBiasDecoder) |
| 89 | { |
| 90 | throw InvalidArgumentException("Bias is enabled but the bias data is invalid"); |
| 91 | } |
| 92 | const armnnUtils::DataLayoutIndexed dataLayoutIndexed(dataLayout); |
| 93 | |
| 94 | const unsigned int channelsIndex = dataLayoutIndexed.GetChannelsIndex(); |
| 95 | const unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 96 | const unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 97 | |
| 98 | unsigned int depthMultiplier = depthwise ? rFilterShape[0] : 1; |
| 99 | unsigned int inputChannels = depthwise ? rFilterShape[1] : rFilterShape[channelsIndex]; |
| 100 | unsigned int outputChannels = depthwise ? inputChannels * depthMultiplier : rFilterShape[0]; |
| 101 | |
| 102 | unsigned int batchSize = rOutputShape[0]; |
| 103 | unsigned int outputHeight = rOutputShape[heightIndex]; |
| 104 | unsigned int outputWidth = rOutputShape[widthIndex]; |
| 105 | unsigned int inputHeight = rInputShape[heightIndex]; |
| 106 | unsigned int inputWidth = rInputShape[widthIndex]; |
| 107 | |
| 108 | unsigned int filterHeight = depthwise ? rFilterShape[2] : rFilterShape[heightIndex]; |
| 109 | unsigned int filterWidth = depthwise ? rFilterShape[3] : rFilterShape[widthIndex]; |
| 110 | |
| 111 | for (unsigned int batchIdx = 0; batchIdx < batchSize; batchIdx++) |
| 112 | { |
| 113 | for (unsigned int cOutput = 0; cOutput < outputChannels; cOutput++) |
| 114 | { |
| 115 | for (unsigned int yOutput = 0; yOutput < outputHeight; yOutput++) |
| 116 | { |
| 117 | for (unsigned int xOutput = 0; xOutput < outputWidth; xOutput++) |
| 118 | { |
| 119 | // This loop goes over each output element. |
| 120 | float sum = 0.0f; |
| 121 | |
| 122 | // For depthwise, each output channel corresponds to exactly one input channel. |
| 123 | // For normal, must loop over each input channel. |
| 124 | for (unsigned int cInput = 0; cInput < (depthwise ? 1 : inputChannels); cInput++) |
| 125 | { |
| 126 | unsigned int depthwiseMultiplierIdx = 0; |
| 127 | if (depthwise) |
| 128 | { |
| 129 | cInput = cOutput / depthMultiplier; |
| 130 | depthwiseMultiplierIdx = cOutput % depthMultiplier; |
| 131 | } |
| 132 | |
| 133 | for (unsigned int yFilter = 0; yFilter < filterHeight; yFilter++) |
| 134 | { |
| 135 | for (unsigned int xFilter = 0; xFilter < filterWidth; xFilter++) |
| 136 | { |
| 137 | // This loop goes over each input element for each output element. |
| 138 | unsigned int filterIndex = 0; |
| 139 | |
| 140 | // Since dimensionality of kernel depends on depthwiseness, so does index. |
| 141 | if (depthwise) |
| 142 | { |
| 143 | filterIndex = depthwiseMultiplierIdx * filterWidth * filterHeight * inputChannels + |
| 144 | cInput * filterWidth * filterHeight + |
| 145 | yFilter * filterWidth + |
| 146 | xFilter; |
| 147 | } |
| 148 | else |
| 149 | { |
Matteo Martincigh | f2aaab3 | 2019-06-06 15:46:22 +0100 | [diff] [blame] | 150 | // Keep this implementation, as using DataLayoutIndexed::GetIndex causes great |
| 151 | // performance regression. |
| 152 | if (dataLayout == DataLayout::NHWC) |
| 153 | { |
| 154 | filterIndex = cOutput * filterHeight * filterWidth * inputChannels + |
| 155 | yFilter * filterWidth * inputChannels + |
| 156 | xFilter * inputChannels + |
| 157 | cInput; |
| 158 | } |
| 159 | else |
| 160 | { |
| 161 | filterIndex = cOutput * filterWidth * filterHeight * inputChannels + |
| 162 | cInput * filterWidth * filterHeight + |
| 163 | yFilter * filterWidth + |
| 164 | xFilter; |
| 165 | } |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 166 | } |
Matteo Martincigh | 18f2d1c | 2019-06-05 13:54:25 +0100 | [diff] [blame] | 167 | |
Aron Virginas-Tar | 5edc881 | 2019-11-05 18:00:21 +0000 | [diff] [blame] | 168 | rFilterDecoder.SetIndex(filterIndex, cOutput); |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 169 | float filterValue = rFilterDecoder.Get(); |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 170 | |
| 171 | unsigned int yInput = yOutput * yStride + yFilter * yDilation; |
| 172 | unsigned int xInput = xOutput * xStride + xFilter * xDilation; |
| 173 | |
| 174 | float inputValue; |
| 175 | |
| 176 | // Check if we're in the padding. |
| 177 | if (yInput < paddingTop || yInput >= inputHeight + paddingTop || |
| 178 | xInput < paddingLeft || xInput >= inputWidth + paddingLeft ) |
| 179 | { |
| 180 | inputValue = 0.0f; |
| 181 | } |
| 182 | else |
| 183 | { |
Matteo Martincigh | f2aaab3 | 2019-06-06 15:46:22 +0100 | [diff] [blame] | 184 | unsigned int inputIndex = 0; |
| 185 | |
| 186 | // Keep this implementation, as using DataLayoutIndexed::GetIndex causes great |
| 187 | // performance regression. |
| 188 | if (dataLayout == DataLayout::NHWC) |
| 189 | { |
| 190 | inputIndex = batchIdx * inputHeight * inputWidth * inputChannels + |
| 191 | (yInput - paddingTop) * inputWidth * inputChannels + |
| 192 | (xInput - paddingLeft) * inputChannels + |
| 193 | cInput; |
| 194 | } |
| 195 | else |
| 196 | { |
| 197 | inputIndex = batchIdx * inputWidth * inputHeight * inputChannels + |
| 198 | inputWidth * inputHeight * cInput + |
| 199 | inputWidth * (yInput - paddingTop) + |
| 200 | xInput - paddingLeft; |
| 201 | } |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 202 | |
Matteo Martincigh | 18f2d1c | 2019-06-05 13:54:25 +0100 | [diff] [blame] | 203 | rInputDecoder[inputIndex]; |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 204 | inputValue = rInputDecoder.Get(); |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 205 | } |
Matteo Martincigh | 18f2d1c | 2019-06-05 13:54:25 +0100 | [diff] [blame] | 206 | |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 207 | sum += filterValue * inputValue; |
| 208 | } |
| 209 | } |
| 210 | } |
| 211 | |
| 212 | if (biasEnabled) |
| 213 | { |
Aron Virginas-Tar | 5edc881 | 2019-11-05 18:00:21 +0000 | [diff] [blame] | 214 | (*pBiasDecoder).SetIndex(cOutput, cOutput); |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 215 | sum += pBiasDecoder->Get(); |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 216 | } |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 217 | |
Matteo Martincigh | 18f2d1c | 2019-06-05 13:54:25 +0100 | [diff] [blame] | 218 | unsigned int outIdx = dataLayoutIndexed.GetIndex(rOutputShape, batchIdx, cOutput, yOutput, xOutput); |
| 219 | |
| 220 | rOutputEncoder[outIdx]; |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 221 | rOutputEncoder.Set(sum); |
Mike Kelly | 9b39832 | 2019-05-22 17:21:49 +0100 | [diff] [blame] | 222 | } |
| 223 | } |
| 224 | } |
| 225 | } |
| 226 | } |
| 227 | |
Aron Virginas-Tar | 5edc881 | 2019-11-05 18:00:21 +0000 | [diff] [blame] | 228 | } // namespace armnn |