FPGA plugin supports only FP16 and FP32 output precision

[RubenPadial]

Hello,

I'm working with Intel FPGA AI Suite Version 2023.1 and OpenVino 2021.4.2 LTS. I have imported a pretrained neural network model and converted to a IR model using model optimizer. It seems that model has been converted properly since I get the same predictions as TensorFlow model.

The command I used is below:

"mo
--saved_model_dir "{path_savedModelPath}"
--input_shape "{lst_inputShape}"
--model_name "{str_modelName}"
--data_type FP16
--output_dir "{path_irTargetPath}""

However, when I try to compile with dla_compiler I get the following error:

"Error occurred.
../compiler/aot_plugin/src/dlia_plugin.cpp:141 FPGA plugin supports only FP16 and FP32 output precision."

Despite using the input parameter "--data_type FP16", .xml file geterated by model optimizer have some layers with precision ="I64" or "I32"

What is the reason? How I can covert the model to an IR model with FP32/16 precision?

Thank you for your help

[Iffa_Intel]

Hi,

could you share your model files for us to verify on our side?

Cordially,

Iffa

[RubenPadial]

Hello @Iffa_Intel 

Find the the following link the files: ConvReshape_NoBatch.zip

[Iffa_Intel]

I checked your model and it seems to be good.

Infer on CPU 

Infer on MYRIAD

MYRIAD required FP16 IR format, so your model should be good to go since it can be inferred as above.

It's recommended to use the latest OpenVINO version (I'm using OV version 2023).

Cordially,

Iffa

[RubenPadial]

Hello @Iffa_Intel 

I Installed OpenVino 2021.4.2 because it is required according to Intel documentation (e.g. https://www.intel.com/content/www/us/en/docs/programmable/768970/2023-1/getting-started-guide.html)

I have updatated OpenVino to 2023.0 version and I get the same error. Could you share with me the script or command you used to infer the model?

[JeanMichelVuillamy]

Ruben,

I would like to confirm that despite what has been mentioned by Iffa_Intel, the Intel FPGA AI Suite version you are using require the use of OpenVino 2021.4.2 as per the documentation.

Jean-Michel Vuillamy (Intel)

[Iffa_Intel]

I verified your model using the OpenVINO benchmark_app: benchmark_app -m path_to_model -d MYRIAD

You could view the command from the screenshot I attached in the previous reply too.

Parsing --data_type FP16 during conversion command is the correct way of getting FP16 precision & from my side, it seems that the model had been correctly converted into IR format with FP16 precision since it can be inferred with MYRIAD which required FP16 IR (as in my previous reply). Your model should be good to go.

Just to clarify:

1. Which hardware/board are you using?
2. Which OS did you use to run with?
3. Did you complete the OpenVINO installation steps as in this guide before proceeding to the FPGA AI Suite? (especially the dependencies & pre-requisite installation part)
4. Another way to test this is to try using IR with FP32 format. By default, if you didn't provide --data type ,it will automatically convert the model into IR with FP32 precision. Any chance that you try this?

Cordially,

Iffa

[RubenPadial]

Hello @Iffa_Intel 

1. Which hardware/board are you using? --> Intel Arria 10 SoC FPGA Dev kit
2. Which OS did you use to run with? --> Ubuntu 20.04
3. Did you complete the OpenVINO installation steps as in this guide before proceeding to the FPGA AI Suite? (especially the dependencies & pre-requisite installation part) --> Yes, it was completed succesfully for Linux. That's the reason I installed OpenVino 2021.4 instead of 2023 version. It is working and IR model was get with this installation using mo function. In addition when I run ". If I try to use OpenVino Runtime toolkit I get an error message requesting OpenVino 2021.4.
   

   dla_compiler --fanalyze-area --march $COREDLA_ROOT/*/A10_Generic.arch​

   I get:
   Exporting input transform to file
   Exporting output transform to file
   Executing area estimate
   Estimated area:
   ALMs: 24099
   ALUTs: 29433
   Registers: 65533
   DSPs: 178
   M20Ks: 611
   Memory ALMs: 1654
4. Another way to test this is to try using IR with FP32 format. By default, if you didn't provide --data type ,it will automatically convert the model into IR with FP32 precision. Any chance that you try this? --> Same result: IR generated succesfully  but same error is get usind dla_compiler.

I'm using the following code

dla_compiler
--march "{path_archPath}"
--network-file "{path_xmlPath}"
--o "{path_binPath}"

Where:

 - path_archPath = $COREDLA_ROOT/example_architectures/A10_Generic.arch

- path_xmlPath = IR .xml filepah

- path_binPath = output directory for compiled model

Is that correct?

I tested bechmark_app as you did and it works on the CPU.

[Iffa_Intel]

Hi,

after some clarification, it is confirmed that:

• Intel® Vision Accelerator Design with an Intel® Arria 10 FPGA (Mustang-F100-A10) speed grade 1 bitstreams and Intel® Programmable Acceleration Card (Intel® PAC) with Intel® Arria® 10 GX FPGA are not included in the Intel Distribution of OpenVINO toolkit 2020.2 release.
• If you use Intel Vision Accelerator Design with an Intel Arria 10 FPGA (Mustang-F100-A10) speed grade 1 and Intel Programmable Acceleration Card (Intel PAC) with Intel Arria 10 GX FPGA, we recommend continuing to use the Intel Distribution of OpenVINO toolkit 2020.1 release.

2020.1 was the last release that supports it which you may refer here.

Cordially,

Iffa

[RubenPadial]

Hello @Iffa_Intel 

Do you refer to this note? https://docs.openvino.ai/2021.4/openvino_docs_install_guides_installing_openvino_linux_fpga.html

But it suggests that OpenVino 2020.3.x will continue supporting Intel Arria 10. it doesn't say anything about discontinuing Intel Arria 10 support from 2020.2.

I have Intel FPGA AI Suite 2023.1 and OpenVino 2021.4.2 running on an Ubuntu 20.04 host PC. I'm working with Intel Arria 10 SoC Dev Kit: https://www.intel.es/content/www/es/es/products/details/fpga/development-kits/arria/10-sx.html. Does it have same requirements /constrains than Intel® Arria 10 FPGA (Mustang-F100-A10) you mentioned?

I don't have noticed about Intel arria 10 incompatibility in FPGA AI suite Documentation. Intel® FPGA AI Suite: Getting Started Guide points that Intel Arria 10 is compatible and there are some examples as well. Intel® FPGA AI Suite: Getting Started Guide also requires OpenVino 2021.4.2.

 Do you mean that any Arria 10 board is no longer supported from OpenVino 2020.1 ahead?  So, do you recommend downgrading to OpenVino 2020.1? And, which FPGA AI suite version is compatible with? 

Thank you for your support.

[JeanMichelVuillamy]

Hello Iffa_Intel,

What you are refering to is an older AI solution running on Intel Arria 10 FPGAs.

In April 2023, Intel released a new solution called the Intel FPGA AI Suite which is far more flexible to address various requirements of the embedded market. Please use this link to access the package and documentation: https://www.intel.com/content/www/us/en/software/programmable/fpga-ai-suite/overview.html

Any questions about this, please let me know.

Jean-Michel Vuillamy (Intel PSG)

[Iffa_Intel]

Hi,

Any version that came after 2020.2 no longer supports Intel® Arria® 10 FPGA and its official (stated in the release notes).

You might be able to install/download/setup, however, issues are expected.

The best is to get the supported and validated version 2020.1 or backward.

You may refer to these older official guides:

1. Intel® Vision Accelerator Design with an Intel® Arria® 10 FPGA
2. Configuration Guide for the Intel® Distribution of OpenVINO™ toolkit 2019R1 and the Intel® Vision Accelerator Design with an Intel® Arria® 10 FPGA (IEI’s Mustang-F100-A10) on Linux*

Cordially,

Iffa

[Iffa_Intel]

Hi,

Intel will no longer monitor this thread since we have provided a solution. If you need any additional information from Intel, please submit a new question. 

Cordially,

Iffa