From 53ef79504b4c881c572735393c2eede5fa556c46 Mon Sep 17 00:00:00 2001
From: Jan Eilers <jan.eilers@arm.com>
Date: Wed, 2 Jun 2021 12:01:25 +0100
Subject: 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
---
 src/backends/backendsCommon/WorkloadData.hpp | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)

(limited to 'src/backends/backendsCommon/WorkloadData.hpp')

diff --git a/src/backends/backendsCommon/WorkloadData.hpp b/src/backends/backendsCommon/WorkloadData.hpp
index 77d4209657..11ce2cb44f 100644
--- a/src/backends/backendsCommon/WorkloadData.hpp
+++ b/src/backends/backendsCommon/WorkloadData.hpp
@@ -208,7 +208,19 @@ struct Convolution2dQueueDescriptor : QueueDescriptorWithParameters<Convolution2
     void Validate(const WorkloadInfo& workloadInfo) const;
 };
 
-// Depthwise Convolution 2D layer workload data.
+/// Depthwise Convolution 2D layer workload data.
+///
+/// @note
+/// The weights are in the format [1, H, W, I*M]. Where I is the input channel size, M the depthwise mutliplier and
+/// H, W is the height and width of the filter kernel. If per channel quantization is applied
+/// the weights will be quantized along the last dimension/axis (I*M) which corresponds to the output channel size.
+/// If per channel quantization is applied the weights tensor will have I*M scales, one for each dimension
+/// of the quantization axis. You have to be aware of this when reshaping the weights tensor.
+/// Splitting the I*M axis, e.g. [1, H, W, I*M] --> [H, W, I, M], won't work without taking care of the
+/// corresponding quantization scales.
+/// If there is no per channel quantization applied reshaping the weights tensor won't cause any issues. There are
+/// preconfigured permutation functions available @link WorkloadUtils.hpp here.
+///
 struct DepthwiseConvolution2dQueueDescriptor : QueueDescriptorWithParameters<DepthwiseConvolution2dDescriptor>
 {
     DepthwiseConvolution2dQueueDescriptor()
-- 
cgit v1.2.1