diff options
author | Jan Eilers <jan.eilers@arm.com> | 2021-04-27 09:21:08 +0100 |
---|---|---|
committer | Jan Eilers <jan.eilers@arm.com> | 2021-06-12 12:08:09 +0100 |
commit | a20d2b8756fb0dea15b1f7620072e510f4977aeb (patch) | |
tree | 00c0a801c7f14b6be1e29f07e85b99a096aac54d | |
parent | 9150bff63a690caa743c471943afe509ebed1044 (diff) | |
download | android-nn-driver-a20d2b8756fb0dea15b1f7620072e510f4977aeb.tar.gz |
IVGCVSW-5826 Change weights layout for depthwise to [1,H,W,I*M]
!armnn:5552
Signed-off-by: Jan Eilers <jan.eilers@arm.com>
Change-Id: I5bbc998f1a29030f718d85b646907c5a314ceecf
-rw-r--r-- | ConversionUtils.hpp | 21 | ||||
-rw-r--r-- | ConversionUtils_1_2.hpp | 18 |
2 files changed, 10 insertions, 29 deletions
diff --git a/ConversionUtils.hpp b/ConversionUtils.hpp index 0f6b783d..f48d58ed 100644 --- a/ConversionUtils.hpp +++ b/ConversionUtils.hpp @@ -472,7 +472,8 @@ void SanitizeBiasQuantizationScale(armnn::TensorInfo& biasInfo, std::transform(biasScales.begin(), biasScales.end(), biasScales.begin(), UpdateBiasScaleValue); biasInfo.SetQuantizationScales(biasScales); - biasInfo.SetQuantizationDim(weightInfo.GetQuantizationDim()); + // bias is expected to be a 1d tensor, set qdim=0 + biasInfo.SetQuantizationDim(0); ALOGV("Bias quantization params have been updated for per-axis quantization"); } @@ -2564,22 +2565,12 @@ bool ConvertDepthwiseConv2d(const HalOperation& operation, const HalModel& model armnn::DepthwiseConvolution2dDescriptor desc; desc.m_DataLayout = armnn::DataLayout::NHWC; - // Reinterpret weight data as [ H, W, I, M ] - armnn::TensorShape weightsShape({ weightsOperand->dimensions[1], - weightsOperand->dimensions[2], - inputInfo.GetShape()[3], - weightsOperand->dimensions[3] / inputInfo.GetShape()[3] }); - - // Swizzle weight data [ H, W, I, M ] -> [ M, I, H, W ] - const armnn::PermutationVector HWIMToMIHW = { 2U, 3U, 1U, 0U }; - + // The layout for weights in depthwise is [ 1, H, W, O] and it's the same in ArmNN. No need to permute anything. const ConstTensorPin weightsPin = ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, model, - data, - HWIMToMIHW, - &weightsShape); + data); // Bias is a 1D tensor const ConstTensorPin biasPin = ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 2, model, data); @@ -2619,8 +2610,8 @@ bool ConvertDepthwiseConv2d(const HalOperation& operation, const HalModel& model return Fail("%s: Operation has invalid inputs", __func__); } - const uint32_t kernelX = weights.GetShape()[3]; - const uint32_t kernelY = weights.GetShape()[2]; + const uint32_t kernelX = weights.GetShape()[2]; + const uint32_t kernelY = weights.GetShape()[1]; const uint32_t inputX = inputInfo.GetShape()[2]; const uint32_t inputY = inputInfo.GetShape()[1]; diff --git a/ConversionUtils_1_2.hpp b/ConversionUtils_1_2.hpp index eec87bf5..f884f7c0 100644 --- a/ConversionUtils_1_2.hpp +++ b/ConversionUtils_1_2.hpp @@ -467,22 +467,12 @@ bool ConvertDepthwiseConv2d_1_2(const HalOperation& operation, const HalModel& m unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); - // Reinterpret weight data as [ H, W, I, M ] - TensorShape weightsShape({ weightsOperand->dimensions[1], - weightsOperand->dimensions[2], - inputInfo.GetShape()[channelsIndex], - weightsOperand->dimensions[3] / inputInfo.GetShape()[channelsIndex] }); - - // Swizzle weight data [ H, W, I, M ] -> [ M, I, H, W ] - const PermutationVector HWIMToMIHW = { 2U, 3U, 1U, 0U }; - + // The layout for weights in depthwise is [ 1, H, W, O] and it's the same in ArmNN. No need to permute anything. const ConstTensorPin weightsPin = ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, model, - data, - HWIMToMIHW, - &weightsShape); + data); // Bias is a 1D tensor const ConstTensorPin biasPin = @@ -516,8 +506,8 @@ bool ConvertDepthwiseConv2d_1_2(const HalOperation& operation, const HalModel& m return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); } - const uint32_t kernelX = weights.GetShape()[3]; - const uint32_t kernelY = weights.GetShape()[2]; + const uint32_t kernelX = weights.GetShape()[2]; + const uint32_t kernelY = weights.GetShape()[1]; const uint32_t inputX = inputInfo.GetShape()[widthIndex]; const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |