[OV CPU] Convert Node converting integer to unsigned one, it did clamp

[bopeng1234]

As Convert Op's Specification said.

Conversion of negative signed integer to unsigned integer value happens in accordance with c++ standard. Notably, result is the unique value of the destination unsigned type that is congruent to the source integer modulo 2^N (where N is the bit width of the destination type). For example, when an int32 value -1 is converted to uint32 the result will be uint32 max which is 4,294,967,295.

Reproduce:

Test when int32 value -1 convert to uint8

• OV GPU/NPU, got 255 (match the spec)
• OV CPU, it uses the boundary to clamp, got 0, not 255 (did't match the spec)

import openvino as ov
import numpy as np

input_data = np.full((2, 2), -1, dtype=np.int32)
expected_output_data = np.full((2, 2), 2**8 - 1, dtype=np.uint8)

input_param = ov.op.Parameter(ov.runtime.Type.i32, ov.Shape([2, 2]))
convert_op = ov.opset1.convert(input_param, destination_type=ov.runtime.Type.u8)
model = ov.Model(results=[convert_op], parameters=[input_param], name="int32_to_uint8_convert")
core = ov.Core()
compiled_model = core.compile_model(model, "CPU")

results = compiled_model(input_data)
actual_output_data = results[compiled_model.outputs[0]]

try:
    print("--- Verification ---")
    np.testing.assert_array_equal(actual_output_data, expected_output_data)
    print("SUCCESS: Actual output matches expected output.")
except AssertionError as e:
    print(f"FAILED: Output mismatch!\n{e}")

--- Verification ---
FAILED: Output mismatch!
Arrays are not equal
...
 ACTUAL: array([[0, 0],[0, 0]], dtype=uint8)
 DESIRED: array([[255, 255],[255, 255]], dtype=uint8)

Tickets:

• CVS-172796

[bopeng1234]

build_jenkins

[bopeng1234]

build_jenkins

[sgbihu]

build_jenkins

[sgbihu]

build_jenkins

[maxnick]

@nshchego, could you please review this PR?

[nshchego]

Why the only one negative value is validated here?

[bopeng1234]

Thanks, changed to a random value.

[nshchego]

Do we really need to check any plugin specific parameters in this test? Looks like we need to compare output blobs only. If so, reusing of the shared tests is sufficient here.

[bopeng1234]

Understand your opilion, if we want to use shared test class, we also need to create a new class to re-write generate_input function with negative value inputs.

root cause is the default test case input value is random generated but with limited range, use int32->uint8 as example, defined in

1. ranges.hpp, it outputs ranges [0, 1000]
2. ranges.cpp, it outputs ranges [0, 255] (cause model's output port is uint8, its range is [0,255])
   then, final range will be [0,255], and results the test input data will be limited in range [0,255], no negative data is tested.

so, if we want to use the shared test, we also need to create a new class to re-write the generate_input func to confirm the negative value (out of range) is tested.

[bopeng1234]

As suggested, moved test from cpu plugin into shared place