iree-org/iree

[TFL] `tfl.lstm` should have a stateful decomposition to other TFLite operations

Open

#9,190 opened on May 23, 2022

View on GitHub
 (3 comments) (0 reactions) (2 assignees)C++ (918 forks)github user discovery
bug 🐞help wantedintegrations/tosa

Repository metrics

Stars
 (3,787 stars)
PR merge metrics
 (PR metrics pending)

Description

Some significant models are blocked by tfl.lstm which can be decomposed into already support tflite operations. Adding the decomposition should unblock these models.

/tmp/test.tflite:0:0: note: see current operation: %95 = "tfl.lstm"(%arg0, %44, %42, %41, %40, %44, %39, %38, %37, %44, %44, %44, %44, %92, %91, %90, %36, %44, %89, %88, %44, %87, %86, %85) ({
}) {asymmetric_quantize_inputs = false, cell_clip = 1.000000e+01 : f32, fused_activation_function = "TANH", kernel_type = "FULL", proj_clip = 0.000000e+00 : f32} : (tensor<1x1280xf32>, none, tensor<2048x1280xi8>, tensor<2048x1280xi8>, tensor<2048x1280xi8>, none, tensor<2048x640xi8>, tensor<2048x640xi8>, tensor<2048x640xi8>, none, none, none, none, tensor<2048xf32>, tensor<2048xf32>, tensor<2048xf32>, tensor<640x2048xi8>, none, tensor<1x640xf32>, tensor<1x2048xf32>, none, tensor<2048xf32>, tensor<2048xf32>, tensor<2048xf32>) -> tensor<*xf32>

Contributor guide