IVGCVSW-6124 ConstTensorsAsInput: Convolution2d
* Add support for Android-nn-driver
!armnn:7382
Signed-off-by: Keith Davis <keith.davis@arm.com>
Signed-off-by: Kevin May <kevin.may@arm.com>
Change-Id: I7ace53cf1c8954c1f2d5588387616d7179ef3bf7
diff --git a/ConversionUtils.hpp b/ConversionUtils.hpp
index c8e7b0d..ca314e2 100644
--- a/ConversionUtils.hpp
+++ b/ConversionUtils.hpp
@@ -1189,7 +1189,8 @@
LayerInputHandle ConvertToLayerInputHandle(const HalOperation& operation,
uint32_t inputIndex,
const HalModel& model,
- ConversionData& data)
+ ConversionData& data,
+ const armnn::PermutationVector& dimensionMappings = g_DontPermute)
{
using HalOperand = typename HalPolicy::Operand;
using HalOperandType = typename HalPolicy::OperandType;
@@ -1252,7 +1253,9 @@
case HalOperandLifeTime::CONSTANT_REFERENCE:
{
// The tensor has an already known constant value, and can be converted into an ArmNN Constant layer.
- ConstTensorPin tensorPin = ConvertOperandToConstTensorPin<HalPolicy>(*operand, model, data);
+ ConstTensorPin tensorPin =
+ ConvertOperandToConstTensorPin<HalPolicy>(*operand, model, data, dimensionMappings);
+
if (tensorPin.IsValid())
{
bool isSupported = false;
@@ -1302,7 +1305,8 @@
LayerInputHandle ConvertToLayerInputHandle(const ::android::hardware::neuralnetworks::V1_3::Operation& operation,
uint32_t inputIndex,
const::android::hardware::neuralnetworks::V1_3::Model& model,
- ConversionData& data)
+ ConversionData& data,
+ const armnn::PermutationVector& dimensionMappings = g_DontPermute)
{
using HalOperand = typename HalPolicy::Operand;
using HalOperandType = typename HalPolicy::OperandType;
@@ -1379,7 +1383,9 @@
case HalOperandLifeTime::CONSTANT_REFERENCE:
{
// The tensor has an already known constant value, and can be converted into an ArmNN Constant layer.
- ConstTensorPin tensorPin = ConvertOperandToConstTensorPin<HalPolicy>(*operand, model, data);
+ ConstTensorPin tensorPin =
+ ConvertOperandToConstTensorPin<HalPolicy>(*operand, model, data, dimensionMappings);
+
if (tensorPin.IsValid())
{
bool isSupported = false;
@@ -2395,18 +2401,21 @@
const armnn::TensorInfo& inputInfo = input.GetTensorInfo();
const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output);
- // ArmNN does not currently support non-fixed weights or bias
- const ConstTensorPin weightsPin = ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, model, data);
- const ConstTensorPin biasPin = ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 2, model, data);
-
- if (!weightsPin.IsValid() || !biasPin.IsValid())
+ LayerInputHandle weightsInput = ConvertToLayerInputHandle<HalPolicy>(operation, 1, model, data);
+ if (!weightsInput.IsValid())
{
return Fail("%s: Operation has invalid inputs", __func__);
}
- armnn::ConstTensor weights = weightsPin.GetConstTensor();
- armnn::ConstTensor bias = biasPin.GetConstTensor();
- SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), inputInfo);
+ LayerInputHandle biasInput = ConvertToLayerInputHandle<HalPolicy>(operation, 2, model, data); // 1D
+ if (!biasInput.IsValid())
+ {
+ return Fail("%s: Operation has invalid inputs", __func__);
+ }
+
+ biasInput.SanitizeQuantizationScale(weightsInput, input);
+ armnn::TensorInfo weightsInfo = weightsInput.GetTensorInfo();
+ armnn::TensorInfo biasInfo = biasInput.GetTensorInfo();
armnn::Convolution2dDescriptor desc;
desc.m_DataLayout = armnn::DataLayout::NHWC;
@@ -2436,8 +2445,8 @@
return Fail("%s: Operation has invalid inputs", __func__);
}
- const uint32_t kernelX = weights.GetShape()[2];
- const uint32_t kernelY = weights.GetShape()[1];
+ const uint32_t kernelX = weightsInfo.GetShape()[2];
+ const uint32_t kernelY = weightsInfo.GetShape()[1];
const uint32_t inputX = inputInfo.GetShape()[2];
const uint32_t inputY = inputInfo.GetShape()[1];
@@ -2450,7 +2459,7 @@
}
desc.m_BiasEnabled = true;
- armnn::Optional<armnn::TensorInfo> biases(bias.GetInfo());
+ armnn::Optional<armnn::TensorInfo> biases(biasInfo);
bool isSupported = false;
auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported)
@@ -2462,7 +2471,7 @@
inputInfo,
outputInfo,
desc,
- weights.GetInfo(),
+ weightsInfo,
biases);
};
@@ -2480,8 +2489,7 @@
return false;
}
- armnn::IConnectableLayer* startLayer =
- data.m_Network->AddConvolution2dLayer(desc, weights, armnn::Optional<armnn::ConstTensor>(bias));
+ armnn::IConnectableLayer* startLayer = data.m_Network->AddConvolution2dLayer(desc);
if (!startLayer)
{
@@ -2490,6 +2498,10 @@
input.Connect(startLayer->GetInputSlot(0));
+ // Connect weights and bias inputs
+ weightsInput.Connect(startLayer->GetInputSlot(1));
+ biasInput.Connect(startLayer->GetInputSlot(2));
+
return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *startLayer, model,
data, nullptr, validateFunc, activation);
}
diff --git a/ConversionUtils_1_2.hpp b/ConversionUtils_1_2.hpp
index b1f2d05..b07d38d 100644
--- a/ConversionUtils_1_2.hpp
+++ b/ConversionUtils_1_2.hpp
@@ -23,6 +23,31 @@
using namespace android::nn;
template<typename HalPolicy,
+ typename HalOperation = typename HalPolicy::Operation,
+ typename HalModel = typename HalPolicy::Model>
+bool IsWeightsValid(const HalOperation& operation,
+ uint32_t inputIndex,
+ const HalModel& model)
+{
+ using HalOperand = typename HalPolicy::Operand;
+ using HalOperandLifeTime = typename HalPolicy::OperandLifeTime;
+ const HalOperand* operand = GetInputOperand<HalPolicy>(operation, inputIndex, model);
+ if (!operand)
+ {
+ Fail("%s: failed to get input operand %i", __func__, inputIndex);
+ return false;
+ }
+
+ if (operand->lifetime != HalOperandLifeTime::CONSTANT_COPY
+ && operand->lifetime != HalOperandLifeTime::CONSTANT_REFERENCE
+ && operand->lifetime != HalOperandLifeTime::NO_VALUE)
+ {
+ return false;
+ }
+ return true;
+}
+
+template<typename HalPolicy,
typename HalOperation = typename HalPolicy::Operation,
typename HalModel = typename HalPolicy::Model>
bool IsQSymmDequantizeForWeights(const HalOperation& operation, const HalModel& model)
@@ -381,26 +406,31 @@
// The NNAPI filter is always OHWI [depth_out, filter_height, filter_width, depth_in] but ArmNN expects the
// filter's height and width indices to match the input's height and width indices so we permute it to OIHW if
// the DataLayout is NCHW
- const ConstTensorPin weightsPin = (desc.m_DataLayout == DataLayout::NCHW) ?
- ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1,
- model, data, OHWIToOIHW) :
- ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 1, model, data);
- const ConstTensorPin biasPin =
- ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 2, model, data);
- if (!weightsPin.IsValid())
+
+ if (!IsWeightsValid<HalPolicy>(operation, 1, model) && desc.m_DataLayout == DataLayout::NCHW)
{
- return Fail("%s: Operation has invalid weights", __func__);
+ return Fail("%s: Operation has unsupported weights HalOperandLifeTime", __func__);
}
- if (!biasPin.IsValid())
+ LayerInputHandle weightsInput = (desc.m_DataLayout == DataLayout::NCHW) ?
+ ConvertToLayerInputHandle<HalPolicy>(operation, 1, model, data, OHWIToOIHW) :
+ ConvertToLayerInputHandle<HalPolicy>(operation, 1, model, data);
+
+ if (!weightsInput.IsValid())
{
- return Fail("%s: Operation has invalid biases", __func__);
+ return Fail("%s: Operation has invalid inputs", __func__);
}
- ConstTensor weights = weightsPin.GetConstTensor();
- ConstTensor bias = biasPin.GetConstTensor();
- SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), inputInfo);
+ LayerInputHandle biasInput = ConvertToLayerInputHandle<HalPolicy>(operation, 2, model, data); // 1D
+ if (!biasInput.IsValid())
+ {
+ return Fail("%s: Operation has invalid inputs", __func__);
+ }
+
+ biasInput.SanitizeQuantizationScale(weightsInput, input);
+ armnn::TensorInfo weightsInfo = weightsInput.GetTensorInfo();
+ armnn::TensorInfo biasInfo = biasInput.GetTensorInfo();
ActivationFn activation;
@@ -419,8 +449,8 @@
armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout);
unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex();
unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex();
- const uint32_t kernelX = weights.GetShape()[widthIndex];
- const uint32_t kernelY = weights.GetShape()[heightIndex];
+ const uint32_t kernelX = weightsInfo.GetShape()[widthIndex];
+ const uint32_t kernelY = weightsInfo.GetShape()[heightIndex];
const uint32_t inputX = inputInfo.GetShape()[widthIndex];
const uint32_t inputY = inputInfo.GetShape()[heightIndex];
@@ -449,7 +479,7 @@
}
desc.m_BiasEnabled = true;
- Optional<TensorInfo> biases(bias.GetInfo());
+ Optional<TensorInfo> biases(biasInfo);
bool isSupported = false;
auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported)
@@ -461,7 +491,7 @@
inputInfo,
outputInfo,
desc,
- weights.GetInfo(),
+ weightsInfo,
biases);
};
@@ -479,8 +509,7 @@
return false;
}
- IConnectableLayer* startLayer =
- data.m_Network->AddConvolution2dLayer(desc, weights, Optional<ConstTensor>(bias));
+ armnn::IConnectableLayer* startLayer = data.m_Network->AddConvolution2dLayer(desc);
if (!startLayer)
{
@@ -488,6 +517,8 @@
}
input.Connect(startLayer->GetInputSlot(0));
+ weightsInput.Connect(startLayer->GetInputSlot(1));
+ biasInput.Connect(startLayer->GetInputSlot(2));
return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *startLayer, model,
data, nullptr, validateFunc, activation);
@@ -1202,8 +1233,11 @@
return false;
}
+ ARMNN_NO_DEPRECATE_WARN_BEGIN
IConnectableLayer* convLayer =
data.m_Network->AddConvolution2dLayer(desc, groupWeights, Optional<ConstTensor>(groupBiases));
+ ARMNN_NO_DEPRECATE_WARN_END
+
if (!convLayer)
{
return Fail("%s: AddConvolution2dLayer failed", __func__);