IVGCVSW-7168 Add Conv2d and Constant support to TOSA Reference Backend
* Added TOSA Conv2d and Constant mappings.
* Added unique naming to mappings based on previous and following
layers, so they are connected correctly.
* Updated existing mappings with new naming convention.
* Added all mappings to one main block in OptimizeSubgraphView.
* Removed isMain from mapping functions.
* Added Conv2d EndToEnd test.
Signed-off-by: Matthew Sloyan <matthew.sloyan@arm.com>
Change-Id: I27c3e238407c32379ce25a1f01dad11523ef5d2b
diff --git a/src/backends/tosaCommon/operatorMappings/Conv2dOperator.cpp b/src/backends/tosaCommon/operatorMappings/Conv2dOperator.cpp
new file mode 100644
index 0000000..9c095d6
--- /dev/null
+++ b/src/backends/tosaCommon/operatorMappings/Conv2dOperator.cpp
@@ -0,0 +1,123 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "Conv2dOperator.hpp"
+
+TosaSerializationBasicBlock* ConvertConv2dToTosaOperator(const Layer* layer,
+ const std::vector<const TensorInfo*>& inputs,
+ const std::vector<const TensorInfo*>& outputs,
+ const Convolution2dDescriptor* conv2dDescriptor)
+{
+ std::vector<std::string> inputNames;
+ std::string outputName = std::string("output0_");
+ std::string blockName = std::string("Op_CONV2D_block_") + GetUniqueTosaMappingID();
+
+ // Set input names for validation purposes only.
+ if(layer == nullptr)
+ {
+ inputNames.emplace_back("input0_");
+ inputNames.emplace_back("input1_");
+ if(conv2dDescriptor->m_BiasEnabled)
+ {
+ inputNames.emplace_back("input2_");
+ }
+ }
+ else
+ {
+ // If a layer is present then the block will be used for execution, so input and output names need to be
+ // determined using the previous and following layers so the graph is connected correctly.
+ // For validation this doesn't matter.
+ for (uint32_t i = 0; i < inputs.size(); ++i)
+ {
+ // Get the layer connected to the input slot and determine unique layer name.
+ Layer& connectedLayer = layer->GetInputSlot(i).GetConnectedOutputSlot()->GetOwningLayer();
+
+ std::string inputName = GenerateUniqueName(connectedLayer, i);
+ inputNames.push_back(inputName);
+ }
+
+ // Get the layer connected to the output slot and determine unique layer name.
+ Layer& connectedLayer = layer->GetOutputSlot().GetConnection(0)->GetOwningLayer();
+
+ outputName = GenerateUniqueName(connectedLayer, 0);
+ }
+
+ std::vector<TosaSerializationTensor*> tensors;
+ std::vector<TosaSerializationOperator*> operators;
+
+ // Setup input Tensor
+ std::vector<int32_t> inputShape0 = GetTosaTensorShape(inputs[0]->GetShape());
+ DType inputDType0 = ArmNNToDType(inputs[0]->GetDataType());
+
+ tensors.push_back(new TosaSerializationTensor(inputNames[0], inputShape0, inputDType0, {}));
+
+ // Only add input tensors if weights and bias are not constant or if running validation.
+ // Constant tensors will be created in the ConvertConstantToTosaOperator function.
+ if(!inputs[1]->IsConstant() || layer == nullptr)
+ {
+ std::vector<int32_t> inputShape1 = GetTosaTensorShape(inputs[1]->GetShape());
+ DType inputDType1 = ArmNNToDType(inputs[1]->GetDataType());
+
+ tensors.push_back(new TosaSerializationTensor(inputNames[1], inputShape1, inputDType1, {}));
+ }
+
+ if(conv2dDescriptor->m_BiasEnabled)
+ {
+ if(!inputs[2]->IsConstant() || layer == nullptr)
+ {
+ std::vector<int32_t> inputShape2 = GetTosaTensorShape(inputs[2]->GetShape());
+ DType inputDType2 = ArmNNToDType(inputs[2]->GetDataType());
+
+ tensors.push_back(new TosaSerializationTensor(inputNames[2], inputShape2, inputDType2, {}));
+ }
+ }
+ else
+ {
+ // If bias is disabled, create a constant bias of 0 as three inputs are required.
+ std::string constantName = std::string("constant_") + GetUniqueTosaMappingID();
+
+ operators.push_back(new TosaSerializationOperator(Op_CONST, Attribute_NONE, nullptr, {}, {constantName}));
+
+ std::vector<uint8_t> uint8Data;
+ std::vector<float> data = { 0.0 };
+
+ TosaSerializationHandler::ConvertF32toU8(data, uint8Data);
+
+ tensors.push_back(new TosaSerializationTensor(constantName, {1}, DType_FP32, uint8Data));
+ inputNames.emplace_back(constantName);
+ }
+
+ // Setup Output Tensor
+ std::vector<int32_t> outputShape0 = GetTosaTensorShape(outputs[0]->GetShape());
+ DType outputDType0 = ArmNNToDType(outputs[0]->GetDataType());
+
+ tensors.push_back(new TosaSerializationTensor(outputName, outputShape0, outputDType0, {}));
+
+ // Set up CONV2D operator
+ std::vector<int> pad = {static_cast<int>(conv2dDescriptor->m_PadTop),
+ static_cast<int>(conv2dDescriptor->m_PadBottom),
+ static_cast<int>(conv2dDescriptor->m_PadLeft),
+ static_cast<int>(conv2dDescriptor->m_PadRight)};
+ std::vector<int> stride = {static_cast<int>(conv2dDescriptor->m_StrideY),
+ static_cast<int>(conv2dDescriptor->m_StrideX)};
+ std::vector<int> dilation = {static_cast<int>(conv2dDescriptor->m_DilationY),
+ static_cast<int>(conv2dDescriptor->m_DilationX)};
+ TosaConvAttribute attribute(pad, dilation, stride, 0, 0, ArmNNToDType(inputs[0]->GetDataType()));
+
+ auto* op = new TosaSerializationOperator(Op_CONV2D,
+ Attribute_ConvAttribute,
+ &attribute,
+ inputNames,
+ {outputName});
+ operators.push_back(op);
+
+ // operatorInputNames/operatorOutputNames ends up being the same as
+ // blockInputNames/blockOutputNames for one-to-one ArmNN to TOSA mappings
+ return new TosaSerializationBasicBlock(blockName, // name
+ operators, // operators
+ tensors, // tensors
+ inputNames, // inputs
+ {outputName}); // outputs
+}
\ No newline at end of file