From e9aa4699ba4c4e16893867dda8dab0d0eed7e7fa Mon Sep 17 00:00:00 2001
From: Matthew Bentham <matthew.bentham@arm.com>
Date: Wed, 24 Apr 2019 14:36:20 +0100
Subject: MLCE-117 Handle more cases for implicit flattening of Fully Connected
 input

Adds new unit test cases, and changes the implementation of
FlattenFullyConnectedInput to more closely match the documentation
of Android NNAPI.

Change-Id: I7ca96b1168b9c7bc78db66f53b0cc776153fd780
Signed-off-by: Matthew Bentham <matthew.bentham@arm.com>
---
 1.0/FullyConnected.hpp             | 14 +++++---------
 test/1.0/FullyConnectedReshape.cpp | 22 ++++++++++++++++++++++
 2 files changed, 27 insertions(+), 9 deletions(-)

diff --git a/1.0/FullyConnected.hpp b/1.0/FullyConnected.hpp
index 0fb029de..26d61e4c 100644
--- a/1.0/FullyConnected.hpp
+++ b/1.0/FullyConnected.hpp
@@ -17,21 +17,17 @@ inline armnn::TensorShape FlattenFullyConnectedInput(const armnn::TensorShape &i
 {
     if (inputShape.GetNumDimensions() > 2U)
     {
-        unsigned int dim0 = inputShape[0];
-        unsigned int dim1 = inputShape[1];
+        unsigned int totalInputElements = inputShape.GetNumElements();
+        unsigned int inputSize = weightsShape[1];
 
-        for (unsigned int i = 2U; i < inputShape.GetNumDimensions(); ++i)
-        {
-            dim1 *= inputShape[i];
-        }
+        unsigned int batchSize = totalInputElements / inputSize;
 
-        unsigned int divisor = weightsShape[1] / dim1;
-        if(dim0 % divisor != 0)
+        if(totalInputElements % batchSize != 0)
         {
             throw std::runtime_error("Failed to deduce tensor shape");
         }
 
-        return armnn::TensorShape({dim0 / divisor, dim1 * divisor});
+        return armnn::TensorShape({batchSize, inputSize});
     }
     else
     {
diff --git a/test/1.0/FullyConnectedReshape.cpp b/test/1.0/FullyConnectedReshape.cpp
index 250f8837..72c90ca5 100644
--- a/test/1.0/FullyConnectedReshape.cpp
+++ b/test/1.0/FullyConnectedReshape.cpp
@@ -13,8 +13,30 @@ BOOST_AUTO_TEST_SUITE(FullyConnectedReshapeTests)
 BOOST_AUTO_TEST_CASE(TestFlattenFullyConnectedInput)
 {
     using armnn::TensorShape;
+
+    // Pass through 2d input
+    BOOST_TEST(FlattenFullyConnectedInput(TensorShape({2,2048}), TensorShape({512, 2048})) ==
+               TensorShape({2, 2048}));
+
+    // Trivial flattening of batched channels
     BOOST_TEST(FlattenFullyConnectedInput(TensorShape({97,1,1,2048}), TensorShape({512, 2048})) ==
                TensorShape({97, 2048}));
+
+    // Flatten single batch of rows
+    BOOST_TEST(FlattenFullyConnectedInput(TensorShape({1,97,1,2048}), TensorShape({512, 2048})) ==
+               TensorShape({97, 2048}));
+
+    // Flatten single batch of columns
+    BOOST_TEST(FlattenFullyConnectedInput(TensorShape({1,1,97,2048}), TensorShape({512, 2048})) ==
+               TensorShape({97, 2048}));
+
+    // Move batches into input dimension
+    BOOST_TEST(FlattenFullyConnectedInput(TensorShape({50,1,1,10}), TensorShape({512, 20})) ==
+               TensorShape({25, 20}));
+
+    // Flatten single batch of 3D data (e.g. convolution output)
+    BOOST_TEST(FlattenFullyConnectedInput(TensorShape({1,16,16,10}), TensorShape({512, 2560})) ==
+               TensorShape({1, 2560}));
 }
 
 BOOST_AUTO_TEST_SUITE_END()
-- 
cgit v1.2.1