Support more Qualcomm SoCs

[jingcheng88]

Support more Qualcomm SoCs when use Qualcomm AI Engine Direc.

[haowhsu-quic]

Hi @jingcheng88 , thank you for using Qualcomm AI Engine Direct. May I know which SoC you are interested?

[jingcheng88]

@haowhsu-quic Qualcomm SA8295. I have tried to add SA8295 myself but failed with following error:

QNN_EXECUTORCH_LOG(
        kLogLevelWarn,
        "[Qnn ExecuTorch] Failed to get soc info for "
        "soc model %d. Using default soc_model=%d",
        htp_options_.soc_model,
        default_soc_model);

One of the places that i modifeid:

typedef enum 
  UNKNOWN_SM = 0,
  SM8450 = 36, // v69
  SM8475 = 42, // v69
  SM8550 = 43, // v73
  SA8295 = 39, // v68
} QcomChipset;

[haowhsu-quic]

Could you try grepping all the file locations including "SM8550" and add SoC information accordingly. From your message, it looks like PopulateSocInfoTable not getting updated.

[jingcheng88]

const std::map<QcomChipset, HtpInfo>& PopulateSocInfoTable() {
  static const std::map<QcomChipset, HtpInfo> soc_info_map{
      {QcomChipset::SM8550,
       {QcomChipset::SM8550,
        QnnHtpDevice_Arch_t::QNN_HTP_DEVICE_ARCH_V73,
        "SM8550",
        8}},
      {QcomChipset::SM8450,
       {QcomChipset::SM8450,
        QnnHtpDevice_Arch_t::QNN_HTP_DEVICE_ARCH_V69,
        "SM8450",
        8}},
      {QcomChipset::SM8475,
       {QcomChipset::SM8475,
        QnnHtpDevice_Arch_t::QNN_HTP_DEVICE_ARCH_V69,
        "SM8475",
        8}},
      {QcomChipset::SA8295,
       {QcomChipset::SA8295,
        QnnHtpDevice_Arch_t::QNN_HTP_DEVICE_ARCH_V68,
        "SA8295",
        8}},
  };
  return soc_info_map;
}

[jingcheng88]

After i recompile libexecutorchdemo.so, the error changed:

W  [Qnn ExecuTorch] Get soc info for soc model 39.  # Add 
W [Qnn ExecuTorch]:  <W> HTP arch will be deprecated, please set SoC id instead.
E [Qnn ExecuTorch]:  <E> Unsupported HTP arch 69 from device custom config
E [Qnn ExecuTorch]:  <E> Failed to parse custom config: 14001
E [Qnn ExecuTorch]:  <E> Failed to parse platform config: 14001

HTP arch should be V68 for SA8295.

[haowhsu-quic]

Yes, SA8295 should be v68. Have you copied the rebuilt pybind libraries (build_x86_64/backends/qualcomm/PyQnn*) to backends/qualcomm/python?
Also, should be aware of pushing the correct skel / stub version to the device.

[kimishpatel]

@jingcheng88 is the error happening at runtime or during delegation?

[jingcheng88]

@kimishpatel Runtime.
@haowhsu-quic I did all those things.

[chiwwang]

Not sure why 69 is passed instead of 68 even with the correct setting in PopulateSocInfoTable. A quick guess might be aarch64 is not re-compiled correctly.

@jingcheng88 Could you please provide a PR or draft PR containing your changes? Best to have reproduce steps and then we can look into this issue.

And for unblocking the progress, you might want to hack a bit to force V68 at

executorch/backends/qualcomm/runtime/backends/htpbackend/HtpDevice.cpp

Line 309 in 9682172

p_custom_config->arch.arch = qcom_target_soc_info_.m_htpArch;

and

executorch/backends/qualcomm/runtime/backends/htpbackend/x86_64/HtpDevicePlatformInfoConfig.cpp

Line 38 in 9682172

p_device_info_extension->onChipDevice.socModel =

.

[jingcheng88]

@chiwwang #1196
https://pytorch.org/executorch/stable/build-run-qualcomm-ai-engine-direct-backend.html
https://pytorch.org/executorch/stable/demo-apps-android.html?highlight=demo

I run the demo app according to the guidelines.

[chiwwang]

Thank you for the information. The PR looks correct. We will give it a try.

BTW, I checked the tutorial https://pytorch.org/executorch/stable/demo-apps-android.html?highlight=demo
One question -- did you re-compile the .pte file and re-push to the device?
I.e., put it to the asset folder of app and trigger gradle clean, gradle build. gradle clean is possibly required since gradle might fail to detect modifications of assets.
ExecuTorch is currently AOT so the dsp_arch info is actually embedded in .pte file.

[jingcheng88]

@chiwwang I delete build_x86, build_android, cmake-out and re-build.
I also re-compile the .pte file, remove old .pte and put new one to the assets.

python -m examples.qualcomm.scripts.deeplab_v3 -b build_android -m SA8295 -c

I will retry again to confirm.

[chiwwang]

Hi @jingcheng88 ,

I fetched your branch, built the code and ran it on a V68 device.

 python -m examples.qualcomm.scripts.deeplab_v3 -b build_android -m SA8295

And it works.

So maybe some problems while the .pte file was packaged into Android demo apk... maybe Android Studio didn't package latest .pte in assets folder?
Could you try to run the above python -m .... to see if the model can work with simple ADB setup? Just remove the -c from the command you used and the script should push related files to the device through adb.