IVGCVSW-6940 ConstTensorsAsInput: DepthwiseConvolution2d - Complete ACL
* Added backend specific optimization & test for CpuAcc and GpuAcc: PermuteDepthwiseConv2dWeights
Signed-off-by: Cathal Corbett <cathal.corbett@arm.com>
Change-Id: I600476b2e9c557a39818a574c1091c9d650b21b1
diff --git a/src/backends/backendsCommon/WorkloadUtils.cpp b/src/backends/backendsCommon/WorkloadUtils.cpp
index d2ae16a..b045530 100644
--- a/src/backends/backendsCommon/WorkloadUtils.cpp
+++ b/src/backends/backendsCommon/WorkloadUtils.cpp
@@ -175,12 +175,14 @@
TensorInfo weightsPermuted;
if (dataLayout == armnn::DataLayout::NHWC)
{
- // No permutation required. Data layouts are the same.
+ // No permutation required. Input and weights data layouts are the same.
aclDepthMultiplier = weightInfo.GetShape()[3] / inputInfo.GetShape()[3];
weightsPermuted = weightInfo;
}
+
else if (dataLayout == armnn::DataLayout::NCHW)
{
+ // Weights permutation required. Weights [N,H,W,C] and input [N,C,H,W] data layouts are different.
// [ 1, H, W, I*M] --> [ 1, I * M, H, W ]
aclDepthMultiplier = weightInfo.GetShape()[3] / inputInfo.GetShape()[1];
PermutationVector permutationVector{ 0, 2, 3, 1 };