TDA2: TIDL-1.1 edma utility

lin rimao

Part Number: TDA2

Hi:

in tidl src code, in the convolution layer, there are edma_utils_scatthergather and edma_utils_autoincrement, and I don't understand why use both of them, here what the difference between them. I refer the edma docs and learn the convolution layer code, but still can't understand them.

over 7 years ago

0 Yordan Kamenov over 7 years ago

TI__Mastermind 42515 points

Hi @lin rimao,

is this the document you referred: PROCESSOR_SDK_VISION_03_05_00_00\ti_components\algorithms\REL.DMAUTILS.00.08.00.02\dmautils\docs\DMAUtilsLibrary_UserGuide.pdf? It has quite detailed explanation of Scater/Gather and Autoincrement mechanisms.

Regards,

Yordan

0 lin rimao over 7 years ago in reply to Yordan Kamenov

Intellectual 370 points

yes, I refered that doc. But in TIDL convolution layer, I can not understand why need both of them.

And now, we have trained a ssd model, current the ssd model can run 15 frames in one sec, is there any solution I can take to run the ssd model 30 frames in one sec through refine the tidl code or something else.

thanks.

0 Yordan Kamenov over 7 years ago in reply to lin rimao

TI__Mastermind 42515 points

Hi,

I have forwarded your question to TIDL experts for further help.

Regards,
Yordan

0 Praveen Eppa1 over 7 years ago in reply to lin rimao

TI__Genius 17580 points

Hi,

TIDL convolution layer is a complex layer implementation in which we had integrated many data flows to support different kernel sizes and different stride and dilation cases and hence used scatter gather DMA when data is spread across different channels ( for example for input and output feature map data transfers), used auto increment when the data is in continues blocks ( for example Bias data).
You can refer to each function header is documented with usage of these DMA implementation functions.

Regarding the other question, did you exploit the sparsity in the model to improve the run time performance.
We have a extension of caffe training framework called caffe-jacinto (github.com/.../caffe-jacinto) which can help to induce sparsity in the model.

Thanks,
Praveen

0 lin rimao over 7 years ago in reply to Praveen Eppa1

Intellectual 370 points

Hi Praveen:

In convolution layer, I noticed for input and output feature map data transfers, both the scatter gather and autoincrement dma utility are used. I don't understand what's the relationship between them.

We have exploit the sparsity used caffe-jacinto, now for 640x320 input model, on one eve the local latency is 180ms, is there any ways (for example through modify the dma tranfer or something in the tidl src code for fixed model) to reduce the latency on one eve.

thanks.

0 lin rimao over 7 years ago in reply to Praveen Eppa1

Intellectual 370 points

Hi,
And for the bias value, if I can use TIDL_edmaCopy transfer the all bias value, but not use autoincrement dma more times, because the bias value num depend numOutCh.