IVGCVSW-2092 Port LSTMCell::Eval to ArmNN

 * Ported Google's LSTM implementation to RefLstmFloat32Workload
 * Fixed the code throughout because of an error in the docs around the
   scratch buffer size
 * Updated IsLstmSupported
 * Added the unit tests

!android-nn-driver:127

Change-Id: I5577b7e39ca52df1a7f102a9b437df6aa99520b6

1 files changed, 4 insertions, 4 deletions

diff --git a/src/backends/cl/workloads/ClLstmFloatWorkload.cpp b/src/backends/cl/workloads/ClLstmFloatWorkload.cpp
index 2a664454e1..f4d8974226 100644
--- a/src/backends/cl/workloads/ClLstmFloatWorkload.cpp
+++ b/src/backends/cl/workloads/ClLstmFloatWorkload.cpp
@@ -116,14 +116,14 @@ ClLstmFloatWorkload::ClLstmFloatWorkload(const LstmQueueDescriptor &descriptor,
     m_ScratchBuffer = std::make_unique<arm_compute::CLTensor>();
     if (m_Data.m_Parameters.m_CifgEnabled)
     {
-        // 2D tensor with dimensions [num_units * 4, batch_size] with CIFG
-        armnn::TensorInfo scratchBuffer1({ batch_size, num_units * 4 }, DataType::Float32);
+        // 2D tensor with dimensions [num_units * 3, batch_size] with CIFG
+        armnn::TensorInfo scratchBuffer1({ batch_size, num_units * 3 }, DataType::Float32);
         BuildArmComputeTensor(*m_ScratchBuffer, scratchBuffer1);
     }
     else
     {
-        // scratch_buffer [num_units * 3, batch_size] without CIFG
-        armnn::TensorInfo scratchBuffer2({ batch_size, num_units * 3 }, DataType::Float32);
+        // scratch_buffer [num_units * 4, batch_size] without CIFG
+        armnn::TensorInfo scratchBuffer2({ batch_size, num_units * 4 }, DataType::Float32);
         BuildArmComputeTensor(*m_ScratchBuffer, scratchBuffer2);
     }