From c833cef6240abb941725a667042b84b936f1e86f Mon Sep 17 00:00:00 2001 From: David Monahan Date: Wed, 3 May 2023 15:53:03 +0100 Subject: IVGCVSW-7612 Implement Split Operators * Added implementation for Split / SplitV in Opaque Delegate Signed-off-by: David Monahan Change-Id: I1ad2925aece7d4e0567e63ecb768348152a70a7a --- delegate/opaque/CMakeLists.txt | 3 +- delegate/opaque/src/Split.hpp | 386 +++++++++++++++++++++++++++++++++ delegate/opaque/src/armnn_delegate.cpp | 12 + 3 files changed, 400 insertions(+), 1 deletion(-) diff --git a/delegate/opaque/CMakeLists.txt b/delegate/opaque/CMakeLists.txt index faeabf1832..a82b75aef8 100644 --- a/delegate/opaque/CMakeLists.txt +++ b/delegate/opaque/CMakeLists.txt @@ -15,7 +15,7 @@ list(APPEND armnnOpaqueDelegateObject_sources src/Comparison.hpp src/Control.hpp src/Convolution.hpp - src/ElementwiseBinary.hpp + src/ElementwiseBinary.hpp src/ElementwiseUnary.hpp src/FullyConnected.hpp src/Gather.hpp @@ -37,6 +37,7 @@ list(APPEND armnnOpaqueDelegateObject_sources src/Slice.hpp src/Softmax.hpp src/SpaceDepth.hpp + src/Split.hpp src/StridedSlice.hpp src/Transpose.hpp src/Unpack.hpp) diff --git a/delegate/opaque/src/Split.hpp b/delegate/opaque/src/Split.hpp index e16969768e..3840ed9620 100644 --- a/delegate/opaque/src/Split.hpp +++ b/delegate/opaque/src/Split.hpp @@ -2,3 +2,389 @@ // Copyright © 2023 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // + +#pragma once + +#include +#include + +#include +#include +#include + +namespace armnnOpaqueDelegate +{ + +constexpr unsigned int MaxNumOfTensorDimensions = 5U; + +TfLiteStatus VisitSplitOperator(DelegateData& delegateData, + TfLiteOpaqueContext* tfLiteContext, + TfLiteOpaqueNode* tfLiteNode, + int nodeIndex, + int32_t tfLiteSplitOperatorCode) +{ + TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 2, nodeIndex)); + + auto* splitParameters = reinterpret_cast(TfLiteOpaqueNodeGetBuiltinData(tfLiteNode)); + int numSplits = NonNegative(splitParameters->num_splits, nodeIndex); + + TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, numSplits, nodeIndex)); + + // Gather input indices and use to get Axis tensor. + const int* inputTensors; + auto numInputs = TfLiteOpaqueNodeNumberOfInputs(tfLiteNode); + if (TfLiteOpaqueNodeInputs(tfLiteNode, &inputTensors, &numInputs) != kTfLiteOk) + { + TF_LITE_OPAQUE_MAYBE_KERNEL_LOG( + tfLiteContext, + "TfLiteArmnnOpaqueDelegate: Unable to gather input tensor indices from node #%d: ", + nodeIndex); + return kTfLiteError; + } + + const TfLiteOpaqueTensor* tfLiteAxisTensor = TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, inputTensors[0]); + if (!IsValid(tfLiteContext, tfLiteAxisTensor, tfLiteSplitOperatorCode, nodeIndex)) + { + return kTfLiteError; + } + + // Use input indices to get input tensor. + const TfLiteOpaqueTensor* tfLiteInputTensor = TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, inputTensors[1]); + if (!IsValid(tfLiteContext, tfLiteInputTensor, tfLiteSplitOperatorCode, nodeIndex)) + { + return kTfLiteError; + } + + // Gather output indices and use to get output tensors. + const int* outputTensors; + if (TfLiteOpaqueNodeOutputs(tfLiteNode, &outputTensors, &numSplits) != kTfLiteOk) + { + TF_LITE_OPAQUE_MAYBE_KERNEL_LOG( + tfLiteContext, + "TfLiteArmnnOpaqueDelegate: Unable to gather output tensor indices from node #%d: ", + nodeIndex); + return kTfLiteError; + } + + const armnn::TensorInfo& inputTensorInfo = GetTensorInfoForTfLiteOpaqueTensor(tfLiteInputTensor); + + ARMNN_ASSERT(GetTensorInfoForTfLiteOpaqueTensor(tfLiteAxisTensor).GetNumElements() == 1); + auto* axisTensorDataPtr = static_cast(TfLiteOpaqueTensorData(tfLiteAxisTensor)); + std::vector axisTensorData(axisTensorDataPtr, axisTensorDataPtr + 1); + int32_t axis = axisTensorData[0]; + + auto inputDimensions = static_cast(inputTensorInfo.GetNumDimensions()); + if (((axis < -inputDimensions) && (axis < 0)) || ((axis >= inputDimensions) && (axis > 0))) + { + // Square bracket denotes inclusive n while parenthesis denotes exclusive n + // E.g. Rank 4 tensor can have axis in range [-4, 3) + // -1 == 3, -2 == 2, -3 == 1, -4 == 0 + TF_LITE_OPAQUE_MAYBE_KERNEL_LOG( + tfLiteContext, + "TfLiteArmnnDelegate: Operation has invalid axis: #%d. Axis must be in range [-n, n) in node #%d:", + axis, nodeIndex); + } + const unsigned int splitDim = ComputeWrappedIndex(axis, inputTensorInfo.GetNumDimensions()); + + std::vector outputs; + for (int i = 0; i < numSplits; ++i) + { + const TfLiteOpaqueTensor* tfLiteOutputTensor = + TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, outputTensors[i]); + if (!IsValid(tfLiteContext, tfLiteOutputTensor, tfLiteSplitOperatorCode, nodeIndex)) + { + return kTfLiteError; + } + outputs.push_back(GetTensorInfoForTfLiteOpaqueTensor(tfLiteOutputTensor, true)); + } + const std::vector> outputTensorInfos(outputs.begin(), outputs.end()); + + auto inputDimSize = inputTensorInfo.GetNumDimensions(); + if (inputDimSize > MaxNumOfTensorDimensions) + { + TF_LITE_OPAQUE_MAYBE_KERNEL_LOG( + tfLiteContext, + "TfLiteArmnnDelegate: The number of dimensions: #%d for input tensors of the split op cannot be " + "greater than #%d in node #%d: ", inputDimSize, MaxNumOfTensorDimensions, nodeIndex); + return kTfLiteError; + } + + std::vector splitterDimSizes(inputDimSize); + + // Add current input shape to splitterDimSizes + for (unsigned int i = 0; i < inputDimSize; ++i) + { + splitterDimSizes[i] = inputTensorInfo.GetShape()[i]; + } + + if (splitterDimSizes[splitDim] % numSplits != 0) + { + TF_LITE_OPAQUE_MAYBE_KERNEL_LOG( + tfLiteContext, + "TfLiteArmnnDelegate: Number of splits #%d must evenly divide the dimension #%d in node #%d: ", + numSplits, splitterDimSizes[splitDim], nodeIndex); + return kTfLiteError; + } + splitterDimSizes[splitDim] /= numSplits; + + armnn::SplitterDescriptor splitDescriptor(numSplits, inputDimSize); + for (int j = 0; j < numSplits; ++j) + { + // Set the size of the views. + for (unsigned int dimIdx = 0; dimIdx < splitterDimSizes.size(); ++dimIdx) + { + splitDescriptor.SetViewSize(j, dimIdx, splitterDimSizes[dimIdx]); + } + splitDescriptor.SetViewOriginCoord(j, splitDim, splitterDimSizes[splitDim] * j); + } + + armnn::BackendId setBackend; + if (!delegateData.m_Network) + { + // Check if supported + bool isSupported = false; + FORWARD_LAYER_OPAQUE_SUPPORT_FUNC("SPLIT", + tfLiteContext, + IsSplitterSupported, + delegateData.m_Backends, + isSupported, + setBackend, + inputTensorInfo, + outputTensorInfos, + splitDescriptor); + return isSupported ? kTfLiteOk : kTfLiteError; + } + + armnn::IConnectableLayer* layer = delegateData.m_Network->AddSplitterLayer(splitDescriptor); + layer->SetBackendId(setBackend); + ARMNN_ASSERT(layer != nullptr); + + for (unsigned int k = 0; k < layer->GetNumOutputSlots(); ++k) + { + layer->GetOutputSlot(k).SetTensorInfo(outputs[k]); + } + + // Connect the input slots + if(Connect(layer, tfLiteContext, tfLiteNode, delegateData) != kTfLiteOk) + { + return kTfLiteError; + } + return kTfLiteOk; +} + +TfLiteStatus VisitSplitVOperator(DelegateData& delegateData, + TfLiteOpaqueContext* tfLiteContext, + TfLiteOpaqueNode* tfLiteNode, + int nodeIndex, + int32_t tfLiteSplitVOperatorCode) +{ + + TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 3, nodeIndex)); + + const int* inputTensors; + auto numInputs = TfLiteOpaqueNodeNumberOfInputs(tfLiteNode); + if (TfLiteOpaqueNodeInputs(tfLiteNode, &inputTensors, &numInputs) != kTfLiteOk) + { + TF_LITE_OPAQUE_MAYBE_KERNEL_LOG( + tfLiteContext, + "TfLiteArmnnOpaqueDelegate: Unable to gather input tensor indices from node #%d: ", + nodeIndex); + return kTfLiteError; + } + + const TfLiteOpaqueTensor* tfLiteInputTensor = TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, inputTensors[0]); + if (!IsValid(tfLiteContext, tfLiteInputTensor, tfLiteSplitVOperatorCode, nodeIndex)) + { + return kTfLiteError; + } + + const TfLiteOpaqueTensor* tfLiteSplitsTensor = TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, inputTensors[1]); + if (!IsValid(tfLiteContext, tfLiteSplitsTensor, tfLiteSplitVOperatorCode, nodeIndex)) + { + return kTfLiteError; + } + + const TfLiteOpaqueTensor* tfLiteAxisTensor = TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, inputTensors[2]); + if (!IsValid(tfLiteContext, tfLiteAxisTensor, tfLiteSplitVOperatorCode, nodeIndex)) + { + return kTfLiteError; + } + + const armnn::TensorInfo& inputTensorInfo = GetTensorInfoForTfLiteOpaqueTensor(tfLiteInputTensor); + const armnn::TensorInfo& splitsTensorInfo = GetTensorInfoForTfLiteOpaqueTensor(tfLiteSplitsTensor); + ARMNN_ASSERT(splitsTensorInfo.GetNumDimensions() == 1); + ARMNN_ASSERT(GetTensorInfoForTfLiteOpaqueTensor(tfLiteAxisTensor).GetNumElements() == 1); + + auto* axisTensorDataPtr = static_cast(TfLiteOpaqueTensorData(tfLiteAxisTensor)); + std::vector axisTensorData(axisTensorDataPtr, axisTensorDataPtr + 1); + int32_t axis = axisTensorData[0]; + + auto inputDimensions = static_cast(inputTensorInfo.GetNumDimensions()); + if (((axis < -inputDimensions) && (axis < 0)) || ((axis >= inputDimensions) && (axis > 0))) + { + TF_LITE_OPAQUE_MAYBE_KERNEL_LOG( + tfLiteContext, + "TfLiteArmnnDelegate: Operation has invalid axis: #%d. Axis must be in range [-n, n) in node #%d:", + axis, nodeIndex); + } + const unsigned int splitDim = ComputeWrappedIndex(axisTensorData[0], inputTensorInfo.GetNumDimensions()); + + auto* splitVParameters = reinterpret_cast(TfLiteOpaqueNodeGetBuiltinData(tfLiteNode)); + int numSplits = 0; + if (splitVParameters) + { + numSplits = NonNegative(splitVParameters->num_splits, nodeIndex); + } + else + { + numSplits = splitsTensorInfo.GetNumElements(); + } + + if (numSplits <= 0) + { + TF_LITE_OPAQUE_MAYBE_KERNEL_LOG( + tfLiteContext, "TfLiteArmnnDelegate: Invalid number of splits %d in node #%d", + numSplits, nodeIndex); + return kTfLiteError; + } + + TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, numSplits, nodeIndex)); + + // Gather output indices and use to get output tensors. + const int* outputTensors; + if (TfLiteOpaqueNodeOutputs(tfLiteNode, &outputTensors, &numSplits) != kTfLiteOk) + { + TF_LITE_OPAQUE_MAYBE_KERNEL_LOG( + tfLiteContext, + "TfLiteArmnnOpaqueDelegate: Unable to gather output tensor indices from node #%d: ", + nodeIndex); + return kTfLiteError; + } + std::vector outputs; + for (int i = 0; i < numSplits; ++i) + { + const TfLiteOpaqueTensor* tfLiteOutputTensor = + TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, outputTensors[i]); + if (!IsValid(tfLiteContext, tfLiteOutputTensor, tfLiteSplitVOperatorCode, nodeIndex)) + { + return kTfLiteError; + } + outputs.push_back(GetTensorInfoForTfLiteOpaqueTensor(tfLiteOutputTensor, true)); + } + const std::vector> outputTensorInfos(outputs.begin(), outputs.end()); + + auto inputDimSize = inputTensorInfo.GetNumDimensions(); + if (inputDimSize > MaxNumOfTensorDimensions) + { + TF_LITE_OPAQUE_MAYBE_KERNEL_LOG( + tfLiteContext, + "TfLiteArmnnDelegate: The number of dimensions: #%d for input tensors of the split op cannot be " + "greater than #%d in node #%d: ", inputDimSize, MaxNumOfTensorDimensions, nodeIndex); + return kTfLiteError; + } + + std::vector splitsTensorData(numSplits); + std::memcpy(splitsTensorData.data(), TfLiteOpaqueTensorData(tfLiteSplitsTensor), splitsTensorInfo.GetNumBytes()); + + + unsigned int index = 0; + unsigned int inferredIndex = 0; + int numberOfInferred = 0; + int splitSum = 0; + + for (auto splitData : splitsTensorData) + { + if (splitData < 0) + { + ++numberOfInferred; + inferredIndex = index; + } + else + { + splitSum += splitData; + } + ++index; + } + + // Check for inferred axis + if (numberOfInferred == 0) + { + if (splitSum != armnn::numeric_cast(inputTensorInfo.GetShape()[splitDim])) + { + TF_LITE_OPAQUE_MAYBE_KERNEL_LOG( + tfLiteContext, "TfLiteArmnnDelegate: SplitV split_sizes does not sum to the dimension of value " + "along split_dim in node #%d", nodeIndex); + return kTfLiteError; + } + } + else if (numberOfInferred == 1) + { + splitsTensorData[inferredIndex] = armnn::numeric_cast(inputTensorInfo.GetShape()[splitDim]) - splitSum; + } + else + { + TF_LITE_OPAQUE_MAYBE_KERNEL_LOG( + tfLiteContext, "TfLiteArmnnDelegate: SplitV cannot infer split size for " + "more than one split in node #%d", + nodeIndex); + return kTfLiteError; + } + + armnn::SplitterDescriptor splitDescriptor(numSplits, inputDimSize); + unsigned int accumSplit = 0; + for (int j = 0; j < numSplits; ++j) + { + unsigned int splitSize = armnn::numeric_cast(splitsTensorData[j]); + + // Set the size of the views. + for (unsigned int dimIdx = 0; dimIdx < inputTensorInfo.GetNumDimensions(); ++dimIdx) + { + unsigned int dimSize = inputTensorInfo.GetShape()[dimIdx]; + if (dimIdx == splitDim) + { + dimSize = splitSize; + } + splitDescriptor.SetViewSize(j, dimIdx, dimSize); + } + + splitDescriptor.SetViewOriginCoord(j, splitDim, accumSplit); + accumSplit += splitSize; + } + + armnn::BackendId setBackend; + if (!delegateData.m_Network) + { + // Check if supported + bool isSupported = false; + FORWARD_LAYER_OPAQUE_SUPPORT_FUNC("SPLITV", + tfLiteContext, + IsSplitterSupported, + delegateData.m_Backends, + isSupported, + setBackend, + inputTensorInfo, + outputTensorInfos, + splitDescriptor); + return isSupported ? kTfLiteOk : kTfLiteError; + } + + armnn::IConnectableLayer* layer = delegateData.m_Network->AddSplitterLayer(splitDescriptor); + layer->SetBackendId(setBackend); + ARMNN_ASSERT(layer != nullptr); + + for (unsigned int k = 0; k < layer->GetNumOutputSlots(); ++k) + { + layer->GetOutputSlot(k).SetTensorInfo(outputs[k]); + } + + // try to connect the Constant Inputs if there are any + if(ProcessInputs(layer,delegateData, tfLiteContext, tfLiteNode) != kTfLiteOk ) + { + return kTfLiteError; + } + + // Connect + return Connect(layer, tfLiteContext, tfLiteNode, delegateData); +} + +} // namespace armnnOpaqueDelegate \ No newline at end of file diff --git a/delegate/opaque/src/armnn_delegate.cpp b/delegate/opaque/src/armnn_delegate.cpp index 2fd8142169..f7476d17c5 100644 --- a/delegate/opaque/src/armnn_delegate.cpp +++ b/delegate/opaque/src/armnn_delegate.cpp @@ -1064,6 +1064,18 @@ TfLiteStatus ArmnnSubgraph::VisitNode(DelegateData& delegateData, tfLiteNode, nodeIndex, kTfLiteBuiltinSpaceToDepth); + case kTfLiteBuiltinSplit: + return VisitSplitOperator(delegateData, + tfLiteContext, + tfLiteNode, + nodeIndex, + kTfLiteBuiltinSplit); + case kTfLiteBuiltinSplitV: + return VisitSplitVOperator(delegateData, + tfLiteContext, + tfLiteNode, + nodeIndex, + kTfLiteBuiltinSplitV); case kTfLiteBuiltinSub: return VisitElementwiseBinaryOperator(delegateData, tfLiteContext, -- cgit v1.2.1