Use DLA directly, without TRT?

Hi,

I’m wondering if I can submit a nvdla loadable to DLA in Xavier, without using TRT?

More precisely, is the kmd/umd for DLA in Xavier compatible with the open-sourced version here?


If not, can I build a the kmd/umd and replace the original ones in Xavier?

Hi,

TensorRT is the only one high-level API for DLA currently.
But you can still try to access the DLA with the low-level driver.

Thanks.

Hi AastaLLL,

Thanks for reply.
I’m now trying with nvdla_runtime with kmd/umd from Xavier,
but I have found some inconsistency between Xavier’s umd and open-sourced version of umd.
Thus I can’t run open-sourced version of nvdla_runtime over Xavier’s umd.

Anywhere I can find header files for Xavier’s umd?

Thanks!

Hi,

May I know the detail error you meet?

Suppose you can compile and execute with the instruction here:
http://nvdla.org/primer.html#compilation-tools-model-creation-and-compilation

Here is the UMD header of NVDLA:
https://github.com/nvdla/sw/tree/master/umd/include

Thanks.

Hi AastaLLL,

I’m trying to run this example
https://github.com/nvdla/sw/blob/master/umd/tests/runtime/RuntimeTest.cpp
on Xavier’s umd and kmd, instead of open-sourced umd and kmd.

And I run into this error:

if (!i->runtime->initEMU())
    ORIGINATE_ERROR(NvDlaError_DeviceNotFound, "runtime->initEMU() failed");

==
But I found that the open-sourced implementation for “bool Runtime::initEMU(void)” returns true if succeed, and false otherwise, while in Xavier’s implemenation, it seems to return “number of available DLA devices”.

That’s why I suspect the Xavier’s umd is inconsistent with open-sourced one.
I think I’ll need the Xavier’s umd headers, not the open-sourced ones, and modify RuntimeTest.cpp to make it work.

Hi,

Thanks for your feedback.
We will try to reproduce this issue on our side and find some suggestion form our internal team.

Stay tuned.

Thanks.

Btw it’ll be good to also have a runable RuntimeTest.cpp for Xavier DLA.

Hi,

After checking with our internal team, the GitHub test app is for OpenDLA and cannot be used on the Xavier.
A possible workaround is to use nvdla_runtime app from OpenDLA with libnvdla_runtime.so released with the Jetpack.

They are API compatible. Would you mind to give it a try?

Thanks.

I don’t quite get it. Did you mean the libnvdla_runtime.so in
JetPack_4.2_Linux_P2888/Linux_for_Tegra/rootfs/usr/lib/aarch64-linux-gnu/tegra/libnvdla_runtime.so ?

Looks like it’s the same as the one already in Xavier’s file system, which is not working, as described before.

Hi,

Sorry about that.

In general, the GitHub source is for OpenDLA and the version of JetPack4.2 is not open source.
But we think they are API compatible. I will check with the developer again.

Really sorry about status missing.

Hi,

Our internal team is going to check this issue.
Will keep updating the progress with you.

Stay tuned.

Hi AastaLLL,

I see nvdla 1.2.0 released, but issue seems not addressed.
Any update on this?

Hi,

Sorry to keep you waiting.

This issue is still checking by our internal team.
Will update more information with you once we got a feedback.

Thanks.

Hi,

We got an update from our internal team.

Current interface are not consistent. We are planning to align the APIs for Xavier DLA and Open DLA in coming up releases.

Thanks.

Are there any updates on the aligning? I need to use the Xavier DLA without TensorRT, please let me know if there is any way to achieve this. Thanks!