IVGCVSW-3223 Fix ref convolution performance regression
authorMatteo Martincigh <matteo.martincigh@arm.com>
Thu, 6 Jun 2019 14:46:22 +0000 (15:46 +0100)
committerTeresa Charlin <teresa.charlinreyes@arm.com>
Fri, 7 Jun 2019 07:54:51 +0000 (07:54 +0000)
commitf2aaab3a06024b5d5c538cc42799fb2c91b4ca2b
treedd83c6927fbc53ab2b235ea85de6ddb3a893f92d
parent99a663140294afd2a4ea91ccc61b7266f735b46a
IVGCVSW-3223 Fix ref convolution performance regression

 * Do not use DataLayoutIndexed::GetIndex for weights and inputs,
   as it causes a large regression in performance
 * It turned out that the calculation of the indexes for the weights
   and inputs was the way it was because of an optimization done
   many months ago
 * Reverted the relevant hunks and added some comments so
   we won't make the same mistake again in the future
 * Made the GetIndex function inline to speed up other usages

Change-Id: I343b2ef0446993086f58b9dea1f0de0ba2d92216
Signed-off-by: Matteo Martincigh <matteo.martincigh@arm.com>
src/armnnUtils/DataLayoutIndexed.cpp
src/armnnUtils/DataLayoutIndexed.hpp
src/backends/reference/workloads/ConvImpl.cpp