Having different calculation result between cpu and npu with nan, inf op.

[WilsonChen0723]

Hi.

I'm having different results for the same calculation with different processing units.

And it will result in precision error in the final.

Result using NPU
-- out_0132__model_dep_encoder_Resize_NPU.npy
(0,1, 8, 0) [ nan, inf, 0.206299, 0.40625,
(0,1,11, 0) [ -inf, nan, 0.136963, 0.171631,

Result using CPU
-- out_0132__model_dep_encoder_Resize_CPU.npy
(0,1, 8, 0) [ 0.313762, 0.260029, 0.206297, 0.406028,
(0,1,11, 0) [ 0.005915, 0.0714487, 0.136982, 0.171628,

Attached files :

- model_input_n_output.zip : models, input npy, the output npy for the CPU, NPU.
- openvino_benchmark_app_src.zip : In order to save the result into npy, we made some changes to the benchmark_app source code（0_before : origin file、1_after : with fix）
- benchmark_app_custom_result_log.txt : The executing log for benchmark_app

Please help, thank you.

[Hairul_Intel]

Hi WilsonChen0723,

Thank you for reaching out to us.

We noticed from your log text file that you're using OpenVINO version 2023.1.0.

As such, please install the latest OpenVINO 2023.2 version through the OpenVINO archives on the download page and see if the issue is resolved.

Furthermore, please do share your device information (CPU, Operating System, etc.) with us.

Regards,

Hairul

[Hairul_Intel]

Hi WilsonChen0723,

Thank you for your question. If you need any additional information from Intel, please submit a new question as this thread is no longer being monitored.

Regards,

Hairul

[WilsonChen0723]

Hi,

CPU : Intel® Core™ Ultra 5 processor 135H

OS : Windows 11 enterprise.

■ 2023.2
-- out_0132__model_dep_encoder_Resize_NPU.npy
(0,1, 8, 0)    [          nan,       -273.75,      0.206299,       0.40625,
(0,1,11, 0)    [          inf,           nan,      0.136963,      0.171631,
-- out_0132__model_dep_encoder_Resize_CPU.npy
(0,1, 8, 0)    [     0.313762,      0.260029,      0.206297,      0.406028,
(0,1,11, 0)    [     0.005915,     0.0714487,      0.136982,      0.171628,

■ 2023.1
-- out_0132__model_dep_encoder_Resize_NPU.npy
(0,1, 8, 0)    [          nan,           inf,      0.206299,       0.40625,
(0,1,11, 0)    [         -inf,           nan,      0.136963,      0.171631,
-- out_0132__model_dep_encoder_Resize_CPU.npy
(0,1, 8, 0)    [     0.313762,      0.260029,      0.206297,      0.406028,
(0,1,11, 0)    [     0.005915,     0.0714487,      0.136982,      0.171628,