[circle-interpreter] Introduce a library for C-Python interaction

[mhs4670go]

This commit introduces a library for C-Python interaction.

Related: #14352
ONE-DCO-1.0-Signed-off-by: seongwoo [email protected]

[mhs4670go]

This is Python side codes.

from cffi import FFI

ffi = FFI()
ffi.cdef("""
  typedef struct InterpreterWrapper InterpreterWrapper;

  InterpreterWrapper *Interpreter_new(const uint8_t *data, const size_t data_size);
  void Interpreter_delete(InterpreterWrapper *intp);
  void Interpreter_interpret(InterpreterWrapper *intp);
  void Interpreter_writeInputTensor(InterpreterWrapper *intp, const int input_idx, const void *data);
  void Interpreter_readOutputTensor(InterpreterWrapper *intp, const int output_idx, void *output, size_t output_size);
""")
C = ffi.dlopen('./libcircle_interpreter_cffi.so')

import torch
class add(torch.nn.Module):
  def __init__(self):
    super().__init__()

  def forward(self, x, y):
    return x + y

  def get_inputs(self):
    return (torch.randn(3), torch.randn(3))

m = add()

import circle_exir
cm = circle_exir.convert(m, m.get_inputs())
model_data = ffi.from_buffer(cm.circle_binary)

import numpy as np

intp = C.Interpreter_new(model_data, len(cm.circle_binary))

inputs = m.get_inputs()
arr1 = inputs[0].numpy()
arr2 = inputs[1].numpy()
c_arr1 = ffi.from_buffer(arr1)
c_arr2 = ffi.from_buffer(arr2)
print (f'Input1: {arr1}')
print (f'Input2: {arr2}')

# Set input
C.Interpreter_writeInputTensor(intp, 0, c_arr1)
C.Interpreter_writeInputTensor(intp, 1, c_arr2)

# Interpret
C.Interpreter_interpret(intp)

# Retrieve output tensor info
from circle_schema import circle
model = circle.Model.Model.GetRootAsModel(cm.circle_binary, 0)
graph = model.Subgraphs(0)
out_tensors = [graph.Tensors(graph.Outputs(o)) for o in range(graph.OutputsLength())]
out_shapes = [t.ShapeAsNumpy() for t in out_tensors]
out_types = [t.Type() for t in out_tensors]

# Get output
result = np.empty(out_shapes[0], dtype=np.float32)
c_result = ffi.from_buffer(result)
C.Interpreter_readOutputTensor(intp, 0, c_result, 12)
print (f'Output: {result}')

# Free memory
C.Interpreter_delete(intp)

python eval.py
Input1: [ 1.5746298  2.3252928 -1.2210928]
Input2: [0.46135813 0.6819443  0.09515246]
Output: [ 2.0359879  3.0072372 -1.1259403]

[seanshpark]

Q) any plans for unit tests ?

[mhs4670go]

Q) any plans for unit tests ?

Please refer to #14376.

[jinevening]

How about throwing an exception if output_size != getTensorSize(output_node) ?

[mhs4670go]

How about adding an exception if output_size != getTensorSize(output_node) ?
Why doesn't writeInputTensor have data_size argument? readOutputTensor has output_size argument.

I'll add input data size argument and add an exception to both sides.

[mhs4670go]

With patched codes, the error message prints like this.

$ python eval.py
C++ Exception: Invalid input size: 12 != 2

[jinevening]

What are these functions for? Is it used in Python-side? If so, could you show the Python codes too?

[mhs4670go]

Here's the python codes.

def check_for_errors():
  error_message = ffi.string(C.get_last_error()).decode('utf-8')
  if error_message:
    C.clear_last_error()
    raise RuntimeError(f'C++ Exception: {error_message}')

# ..

try:
  C.Interpreter_writeInputTensor(intp, 0, c_arr1, 123456)
  check_for_errors()
except RuntimeError as e:
  print(e)

[jinevening]

LGTM

The interface seems a bit tricky to use, but I guess it may be wrapped by other functions in python.

[seanshpark]

LGTM