IVGCVSW-5826 Change weights layout for depthwise to [1,H,W,I*M]
* This change is necessary because tflite uses a [1,H,W,I*M] format
and uses the I*M dimension for per axis quantization. Our previous
layout [M,I,H,W] can't handle the correlating quantization scales.
* Updates Onnx-, TfLiteParser and TfliteDelegate
* Updates the CpuRef, CpuAcc and GpuAcc backends
* Adjusts unit tests
* Adds test to ensure models with old layout can still be read and
executed
* Adds conversion function to previous layout [1,H,W,I*M] --> [M,I,H,W]
which can be used by backend developers
!android-nn-driver:5553
Signed-off-by: Jan Eilers <jan.eilers@arm.com>
Change-Id: Ifef23368b8c3702cf315a5838d214f7dc13c0152
diff --git a/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp b/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp
index 50cdb0a..9a9977b 100644
--- a/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp
+++ b/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp
@@ -33,12 +33,11 @@
const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
- // ArmNN's weight format is [ M, I, H, W ]
- const unsigned int aclDepthMultiplier = weights.GetShape()[0];
-
- // Convert the weight format from ArmNN's [ M, I, H, W ] (does NOT depend on the data layout) to either
- // [ 1, H, W, I * M ] (if NHWC) or [ 1, I * M, H, W ] (if NCHW), as required by the compute library
- TensorInfo weightsPermuted = ConvertWeightTensorInfoFromArmnnToAcl(weights, descriptor.m_DataLayout);
+ // ArmNN's weight format is usually [ M, I, H, W ] but for depthwise its [ 1, H, W, I*M]
+ // Permute to [ 1, I * M, H, W ] (if NCHW) as required by the compute library
+ unsigned int aclDepthMultiplier;
+ TensorInfo weightsPermuted;
+ std::tie(weightsPermuted, aclDepthMultiplier) = Convert1HWOTensorInfoToAcl(weights, input,descriptor.m_DataLayout);
// Convert the weights into the compute library format
const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weightsPermuted, descriptor.m_DataLayout);
@@ -79,14 +78,15 @@
const arm_compute::CLCompileContext& clCompileContext)
: BaseWorkload<DepthwiseConvolution2dQueueDescriptor>(descriptor, info)
{
- // Allocate a buffer for the swizzling of the weight tensor
+ // ArmNN's weight format is usually [ M, I, H, W ] but for depthwise its [ 1, H, W, I*M]
+ // Permute to [ 1, I * M, H, W ] (if NCHW), as required by the compute library
+ ConstTensor weightPermuted;
+ unsigned int depthMultiplier;
std::unique_ptr<unsigned char[]> permuteBuffer(new unsigned char[m_Data.m_Weight->GetTensorInfo().GetNumBytes()]);
-
- // Convert the weight format from ArmNN's [ M, I, H, W ] (does NOT depend on the data layout) to either
- // [ 1, H, W, I * M ] (if NHWC) or [ 1, I * M, H, W ] (if NCHW), as required by the compute library
- ConstTensor weightPermuted = ConvertWeightTensorFromArmnnToAcl(m_Data.m_Weight,
- m_Data.m_Parameters.m_DataLayout,
- permuteBuffer.get());
+ std::tie(weightPermuted, depthMultiplier) = Convert1HWOTensorToAcl(m_Data.m_Weight,
+ info.m_InputTensorInfos[0],
+ m_Data.m_Parameters.m_DataLayout,
+ permuteBuffer.get());
// Convert the weights into the compute library format
m_KernelTensor = std::make_unique<arm_compute::CLTensor>();
@@ -113,12 +113,6 @@
input.info()->set_data_layout(aclDataLayout);
output.info()->set_data_layout(aclDataLayout);
- // ArmNN's weight format is [ M, I, H, W ]
- auto& weightInfo = m_Data.m_Weight->GetTensorInfo();
-
- // Get the depth multiplier
- const unsigned int depthMultiplier = weightInfo.GetShape()[0];
-
arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters);
const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);