TMDSEVM572X: TIDL jDetNet on single DSP

Hi Niklas,

Can you try setting the showHeapStats flag?  Check here for example output:

http://downloads.ti.com/mctools/esd/docs/tidl-api/using_api.html#sizing-device-side-heaps

Regards,
Mike

Hi Mike,

thanks for the link! I tried as outlined, with the following result:

config = tidl.Configuration()
config.read_from_file(configFile)
config.show_heap_stats = True
config.param_heap_size = 2972496 # derived from TIDL output
config.network_heap_size = 128*1024**2 # still a guess

config.layer_index_to_layer_group_id = {i:1 for i in range(43)}
dsp = tidl.Executor(tidl.DeviceType.DSP, set([tidl.DeviceId.ID0]),config,1)

Output:

dsp = tidl.Executor(tidl.DeviceType.DSP, set([tidl.DeviceId.ID0]),config,1)
Traceback (most recent call last):

  File "<ipython-input-10-7b50316b580f>", line 1, in <module>
    dsp = tidl.Executor(tidl.DeviceType.DSP, set([tidl.DeviceId.ID0]),config,1)

TidlError: TIDL Error: [src/execution_object.cpp, Wait, 617]: Memory allocation failed on device

[eve 0]         TIDL Device Trace: PARAM heap: Size 2972496, Free 0, Total requested 2972496

Note that there is no message concerning the NETWORK heap size, as seen in the api doc. When trying a smaller network, both messages appear correctly.

Any ideas?

Bests,

Niklas

Any updates regarding this issue?

Hi Niklas,

Thank you for your patience - the engineering that normally answers TIDL questions on E2E is on vacation.

I have reached out to the TIDL development team on your issue.

In the meantime, can verify the CMEM driver is loading properly?  Can you respond back with the output from this command:

cat /proc/cmem

Regards,
Mike

Hi Mike,

thanks for your response.

Here is the output:

root@am57xx-evm:~# cat /proc/cmem

Block 0: Pool 0: 1 bufs size 0x20000000 (0x20000000 requested)

Pool 0 busy bufs:
id 0: phys addr 0xa0000000 (cached)

Pool 0 free bufs:

In this case, I am running with 512mb CMEM.

Bests,

Niklas

Hi Niklas,

Our TIDL forum expert will be back in the office on Monday and should be able to get to the bottom of your issue.

In the meantime, check out this link: https://downloads.ti.com/mctools/esd/docs/tidl-api/using_api.html?#sizing-device-side-heaps

There are several examples that come with the TI Linux SDK shown.  If you pass the "-v" option, you will get a verbose output from the demo app, showing the memory allocation.  It would be a good test to try one of the known working demo applications with your CMEM configuration to determine if there is something missing in your Python code, or there is a CMEM configuration issue.

Regards,
Mike

Hi Mike,

I tried several examples:

root@am57xx-evm:/usr/share/ti/tidl/examples/imagenet# ./imagenet -d 1 -e 0
Input: ../test/testvecs/input/objects/cat-pet-animal-domestic-104827.jpeg
1: tabby,   prob = 52.55%
2: Egyptian_cat,   prob = 21.18%
3: tiger_cat,   prob = 17.65%
Loop total time (including read/write/opencv/print/etc):  209.6ms
imagenet PASSED

./mnist -d 1 -e 0
Input images: ../test/testvecs/input/digits10_images_28x28.y
Input labels: ../test/testvecs/input/digits10_labels_10x1.y
0
1
2
3
4
5
6
7
8
9
Device total time:  45.05ms
Loop total time (including read/write/print/etc):  46.19ms
Accuracy:    100%
mnist PASSED

Note, that I explicitly set the amount of DSPs to 1 and EVEs to 0 to verify that a single DSP is able to run the sample networks.

The "ssd_multibox" example works as well, but not out of the box on DSP=1 and EVE=0.

However, the "classification" example fails:

root@am57xx-evm:/usr/share/ti/tidl/examples/classification# ./tidl_classification   
tidl FAILED
root@am57xx-evm:/usr/share/ti/tidl/examples/classification# ./tidl_classification -d 2 -e 2
tidl FAILED
root@am57xx-evm:/usr/share/ti/tidl/examples/classification# ./tidl_classification -v       
tidl FAILED

Bests

Niklas

Hi Niklas,

Which version of Processor SDK Linux are you using? There was bug in earlier versions of the SDK for which the JDetNet model was failing to run on DSP alone. This issue was addressed in Processor SDK version 6.0. Let me know if you are saying that SDK 6.0 is also failing to run the model on DSP alone. 

Regards,

Manisha

Hi Manisha,

thanks for your reply. We are using SDK v. 6.0.

Other networks, like the mobilenet 0.5 detection framework work just fine.

Bests

Niklas