diff options
author | Matteo Martincigh <matteo.martincigh@arm.com> | 2018-11-14 12:39:55 +0000 |
---|---|---|
committer | Matteo Martincigh <matteo.martincigh@arm.com> | 2018-11-16 09:05:24 +0000 |
commit | a65b7aeafc0ef6acf40e4a8a6d36206bf53d717c (patch) | |
tree | d62257a911f3a4a4ed99243d4860a2453e95ec98 /src/backends/neon | |
parent | 74ba3dc7113e51cf11ab772ee1eb030c07a7dda5 (diff) | |
download | armnn-a65b7aeafc0ef6acf40e4a8a6d36206bf53d717c.tar.gz |
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
Diffstat (limited to 'src/backends/neon')
-rw-r--r-- | src/backends/neon/workloads/NeonLstmFloatWorkload.cpp | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/src/backends/neon/workloads/NeonLstmFloatWorkload.cpp b/src/backends/neon/workloads/NeonLstmFloatWorkload.cpp index d03454b705..1ab269ff56 100644 --- a/src/backends/neon/workloads/NeonLstmFloatWorkload.cpp +++ b/src/backends/neon/workloads/NeonLstmFloatWorkload.cpp @@ -114,13 +114,13 @@ NeonLstmFloatWorkload::NeonLstmFloatWorkload(const LstmQueueDescriptor &descript 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); + 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); + armnn::TensorInfo scratchBuffer2({ batch_size, num_units * 4 }, DataType::Float32); BuildArmComputeTensor(*m_ScratchBuffer, scratchBuffer2); } |