diff options
| author | Francis Murtagh <francis.murtagh@arm.com> | 2022-07-22 10:23:41 +0100 |
|---|---|---|
| committer | Francis Murtagh <francis.murtagh@arm.com> | 2022-07-22 10:23:41 +0100 |
| commit | 89cc6b39376419565505c59c693a107a5edd47aa (patch) | |
| tree | 657123d837744e982cc87c4ef024a5fbb60a0c8a /src/armnn/Network.cpp | |
| parent | 5b1f53975d7fa482475af781ffa2a2277f40e0c9 (diff) | |
| download | armnn-89cc6b39376419565505c59c693a107a5edd47aa.tar.gz | |
IVGCVSW-6978: RedirectMembersToConstantInputs does not work with Fp32NetworkToBf16Converter
* Fuse FP32ToBF16Layers with Constant Layer so Conv2d/FullyConnected
can have their weights redirected.
* If BF16 Unsupported in Conv2d || FullyConnected revert fused
Constant Layer to FP32
Change-Id: If523c708a822659d64597d9ae39cca1c2f84b76f
Signed-off-by: Francis Murtagh <francis.murtagh@arm.com>
Diffstat (limited to 'src/armnn/Network.cpp')
| -rw-r--r-- | src/armnn/Network.cpp | 17 |
1 files changed, 12 insertions, 5 deletions
diff --git a/src/armnn/Network.cpp b/src/armnn/Network.cpp index 8fe4445dcf..5d443068ce 100644 --- a/src/armnn/Network.cpp +++ b/src/armnn/Network.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -790,13 +790,18 @@ OptimizationResult AttemptBackendAssignment(BackendSettings& backendSettings, } else if (dataTypeIn == DataType::BFloat16 || dataTypeOut == DataType::BFloat16) { + const auto layerType = layer->GetType(); if (IWorkloadFactory::IsLayerSupported(*layer, DataType::Float32, reasonIfUnsupported) - && layer->GetType() != LayerType::ConvertFp32ToBf16 - && layer->GetType() != LayerType::ConvertBf16ToFp32) + && layerType != LayerType::ConvertFp32ToBf16 + && layerType != LayerType::ConvertBf16ToFp32) { - // Insert BF16 -> FP32 conversion layer before current layer + bool revertConstantWeightsConversion = RevertConstantWeightsToFP32(layer); + + // Insert BF16 -> FP32 conversion layer before current layer. + // Unless we have reverted Constant Weights Type above. std::vector<ConvertBf16ToFp32Layer*> convertBf16ToFp32Layers; - if (dataTypeIn == DataType::BFloat16) + if (dataTypeIn == DataType::BFloat16 && dataTypeOut != DataType::BFloat16 + && !revertConstantWeightsConversion) { convertBf16ToFp32Layers = InsertConvertBf16ToFp32LayersBefore(graph, *layer); @@ -1759,10 +1764,12 @@ IOptimizedNetworkPtr Optimize(const Graph& inGraph, // If Fp32 to Bf16 optimization is set convert Fp32 network to Bf16 // Convert input of Convolution2d and FullyConnected from Fp32 to Bf16 // Only Constant weight of Convolution2d and FullyConnected are converted from Fp32 to Bf16 + // Constant and Fp32ToBf16 layers will also be fused so conversion is no longer needed at inference time if (options.m_ReduceFp32ToBf16) { ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "Optimizer_ReduceFp32ToBf16"); Optimizer::Pass(optGraph, MakeOptimizations(Fp32NetworkToBf16Converter())); + Optimizer::Pass(optGraph, MakeOptimizations(FuseConversionLayersIntoConstLayers())); } // Initialize backend settings |