From 0ac0804e76e098695ee2b8a9e24e2f0a1efc324f Mon Sep 17 00:00:00 2001 From: Fredrik Svedberg Date: Tue, 11 Apr 2023 22:35:04 +0200 Subject: MLBEDSW-7196 Add LSTM support Added int8 and int16 UNIDIRECTIONAL_SEQUENCE_LSTM support. The implementation does not include support for: * CIFG * Peephole * Projection * Normalisation This change also: * Removed unused Op.BlockLSTM operation type. * Removed the only one consumer limitation on putting the SplitSliceRead on the tensor consumer(s), if all consumers fullfills the requirements * Added Op.VariableTensorWrite as a Operation.memory_function to make sure writes to variable tensors: * Always use linear mode * Are not moved to fast scratch * Are not fused with other elementwise operation tensor ranges Change-Id: Ief831738924ac3d1f2ba6d41f10bd6dc969911f3 Signed-off-by: Fredrik Svedberg --- SUPPORTED_OPS.md | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) (limited to 'SUPPORTED_OPS.md') diff --git a/SUPPORTED_OPS.md b/SUPPORTED_OPS.md index 08c63e7c..f641d3f2 100644 --- a/SUPPORTED_OPS.md +++ b/SUPPORTED_OPS.md @@ -55,6 +55,7 @@ Please check the supported operator list for your chosen runtime for further inf | SUB | [Generic](#tflite-generic-constraints), [Specific](#tflite-sub-constraints) | | TANH | [Generic](#tflite-generic-constraints) | | TRANSPOSE_CONV | [Generic](#tflite-generic-constraints), [Specific](#tflite-transpose_conv-constraints) | +| UNIDIRECTIONAL_SEQUENCE_LSTM | [Generic](#tflite-generic-constraints), [Specific](#tflite-unidirectional_sequence_lstm-constraints) | | UNPACK | [Generic](#tflite-generic-constraints) | ### TFLite Generic Constraints @@ -356,3 +357,19 @@ This is a list of constraints that the TRANSPOSE_CONV operator must satisfy in o - SAME padding: OFM dimensions must equal IFM dimensions multiplied by stride - VALID padding: OFM dimensions must equal IFM dimensions multiplied by stride, minus difference between kernel size and stride + +### TFLite UNIDIRECTIONAL_SEQUENCE_LSTM Constraints + +This is a list of constraints that the UNIDIRECTIONAL_SEQUENCE_LSTM operator must satisfy in order to be scheduled on the NPU. + +- IFM must be int8 or int16 +- IFM and OFM data types must match +- IFM and OFM must have 3D shape +- Must have 24 input tensors +- Must have 5 intermediate tensors +- State tensors must be variable +- Must not use CIFG +- Must not use Peephole +- Must not use Projection +- Must not use Normalisation +- All input and recurrent weights must be available -- cgit v1.2.1