[TDA4M] Why is there an error when I use a bigger image for input?

Hi Yongsig,

    Can you set debugTraceLevel = 2 in infer config file and share the log after running it on EVM.

Regards,

Anshu

Hi anshu. I always thank you.

First, It's log on PC. I think It's OK. But there is a weird message as below.

Starting Layer # - 76
Starting Layer # - 77
Processing Layer # - 77
OUT_FEAT_MAP READ MSMC not enough in tidl_conv2d_mma
End of Layer # - 77
42 13.37301 -5.03790 5.41418 1
Starting Layer # - 78
TSC Mega Cycles = 37958.91 ...testvecs/input/img_640x352.bmp
testvecs/input/img_640x352.bmp
A : 0, 0.0000, 0.0000, 324

The layer 77 is covolution which has 3x3 kernel, pad 1 and stride 1. I configured MSMC size to 7936 KB when I used TI import tool. Why is a message printed like that?

And, It's log on target. It's the same network as above one. You know a different thing is only image size.

Processing Cnt : 0, InstCnt : 0 ----------------------- TIDL Process with TARGET DATA FLOW ------------------------

# 0testvecs/input/img_1280x704.bmp
. ..Starting Layer # - 0
Starting Layer # - 1
0 1.00000 0.00000 255.00000 0
Processing Layer # - 1
End of Layer # - 1
1 29.40893 0.00000 6.82669 0
Starting Layer # - 2
Processing Layer # - 2
End of Layer # - 2
2 19.95841 0.00000 9.61472 0
Starting Layer # - 3
Processing Layer # - 3
End of Layer # - 3
4 10.60676 -6.45318 12.06538 1
Starting Layer # - 4
Processing Layer # - 4
End of Layer # - 4
3 10.01728 -7.15482 11.35820 1
Starting Layer # - 5
Processing Layer # - 5
Maximum number of splits for a big transfer is more than supported
End of Layer # - 5

..... ..... .....

Starting Layer # - 76
Starting Layer # - 77
Processing Layer # - 77
End of Layer # - 77
42 9.63938 -7.16487 8.05093 1
Starting Layer # - 78
TSC Mega Cycles = 47514.70 Error at line: 351 : in file src/tidl_tb.c, of function : tidl_tb_algProcess
Invalid Error Type!
Error at line: 558 : in file src/tidl_tb.c, of function : tidlMultiInstanceTest
Invalid Error Type!
Error at line: 564 : in file src/tidl_tb.c, of function : tidlMultiInstanceTest
Invalid Error Type!

We can see 'Maximum number of splits for a big transfer is more than supported' on layer5. The layer 5 is convolution which has 3x3 kernel, pad 1 and stride 1. The width is 352 and the height is 640 as well.

Why is there an error like that?

Best regards

Yongsig

Hi Yongsig,

    We will need a bit more information to understand why this issue is coming. Can you share the following file :

ti_dl/utils/perfsim/<your import config file>/bufinfolog.txt

  If you cannot share the complete file then can you share the entry corresponding to the layer where you are getting "Maximum number of splits for a big transfer is more than supported" message. I am giving you a sample information for one of the layer :

------Layer #(Type) [Exec ID , Data ID] --[Ni x inW x inH] => [No x outW x outH] [Ni/G] [dataflowType] [preFetch, preFetchAlign, procSize, inPlaneSize] [dmaFreq] [dmaFreqWt] [kernelFreq] [In Data Ids] -----

----- 4( Conv) [3, 4] --[12 x 128 x 56] => [24 x 64 x 28] *** [6] ***[ROW_L] ***[130, 258, 7225, 7483]**** [1], [1],[1] -[2 ]---
MSMC, DMA, 1d80( 7552), 1d3b( 7483), c( 12), 16280( 90752), 0, 0 |||| L2, DMA, 1e40( 7744), 1e40( 7744), 6( 6), b480( 46208), 0, 0
NONE, NA, 0( 0), 0( 0), 0( 0), 0( 0), 0, ffffffff ||||MSMC, CPU, 880( 2176), 79f( 1951), 18( 24), cc80( 52352), 42, 2fb3e
DDR, DMA, 37( 55), 37( 55), 18( 24), 580( 1408), 0, 12380 |||| L2, DMA, 37( 55), 37( 55), c( 12), 300( 768), 0, b480
MSMC, CPU, 80( 128), 0( 0), 0( 0), c00( 3072), 0, 16280

Exec ID  is the field which should be same as the layer number where you are seeing the issue.

Regards,

Anshu

First of all, I'm so sorry, I can't share details that you wanted. 

But I found the difference thing on preFetch. It has minus value. Is it OK?

Best regards

Yongsig.

Yongsig,

   I can understand that you cannot share the full log. But can you share the log of the layer where the issue is coming ( as mentioned in my previous response)?

Regards,

Anshu

We can't share any channel and input information of the layer. Even if It's the only one.....I'm so sorry.

So, It's hard to know the reason using only preFetch value, isn't it?

Best regards

Yongsig.

Hi Yongsig,

   Negative value of pre-fetch is not expected. But its difficult to comment much without getting more details. Can you create a dummy network with some other channel numbers and but other kernel property being same and see if you are able to re-produce the issue. If yes will you be able to share this dummy network?

Regards,

Anshu

Hi Anshu.

Thank you for your reply.

I'm trying to make a dummy network to share that network which has the same issue.

I think It'll take a couple of days for me to share that network.

Best regards

Yongsig. 

Hi Yongsig,

Do you have a dummy network?

Regards,

Rishabh

Hi, Rishabh

I tried to make a dummy network which has the same issue, but I couldn't.

So, My team are thinking how to share our information using like webex. But we don't decide that yet. The decision will be done soon.

If we decide that, I'll contact again here.

Best regards

Yongsig.

Hi Yongsig,

Please let us know when you can share the dummy network.

Regards,

Rishabh

Hi Rishabh

I tried a lot of times to make a dummy network which has the same issue.

I need around 2 weeks to make that.

Best regards

Yongsig.