RTOS/TDA2P-ACD: ORB performance on EVE : Estimated DDR BW

Sujith

Genius 12455 points

Part Number: TDA2P-ACD

Tool/software: TI-RTOS

Hi Anshu, Pramod,

Can you please provide us with expected DDR load for ORB with following parameters.

Data Format RGBA 8888, 4 video frame, @ 60 FPS.

Scales are 256x540 & 197x415

Following are some assumptions

1. that data is in DDR and internal memory

2. Scales generated by independent scaler

3. Scales generation is done in EVE

Regards,Sujtih

over 7 years ago

0 Anshu Jain over 7 years ago

TI__Guru 56820 points

Hi Sujith,

As per the following assumption the details of approximate estimated performance is shown below :

Parameters assumed: 256x540 x4 + 197x415 x4 : 4 image regions of resolution 256x540 + 4 image regions of 197x415. For each image region the 2 scale pyramids are created and ORB features are computed. The final output of this processing is 256 bit ORB descriptors and total 512 descriptors

0 Anshu Jain over 7 years ago in reply to Anshu Jain

TI__Guru 56820 points

Hi Sujith,

Please find attached the excel sheet used for calucation. This has some small correction ( Which will not have much impact overall in the calculation shared earlier).

orb estimation.xlsx

0 Anshu Jain over 7 years ago in reply to Anshu Jain

TI__Guru 56820 points

Hi Sujith,

I have updated the performance and DDR BW excel sheet based on the discussion we had.

4786.orb estimation.xlsx

0 Pramod Kumar Swami over 7 years ago in reply to Anshu Jain

TI__Genius 14000 points

One clarification : The input at ORB is assumed as grey scale single 8-bit channel.

With Regards,

Pramod

0 Sujith over 7 years ago in reply to Pramod Kumar Swami

TI__Genius 12455 points

Thanks Anshu, Pramod,

The DDR write bandwidth for the "rBrief" is not included, any reasons?

Also, when operating with ARGB8888, additional cycles required to convert to YUV/Luma

Regards,
Sujith

0 Anshu Jain over 7 years ago in reply to Sujith

TI__Guru 56820 points

Hi Sujith,

PFA the corrected performance and DDR BW excel sheet.

Regards,

Anshu

8424.orb estimation.xlsx

0 Koichi Saito over 7 years ago in reply to Anshu Jain

TI__Intellectual 2525 points

Hi Anshu, Pramod,

I got feedback from customer after their review.

(1) Feature matching should be modified to 2.38 cycles ? Because feature matching in eve_sw_datasheet.pdf takes 2.18 cycles for vcop_feature_matching and 2.38 cyles for Applet feature matching.

(2) Customer needs bi-directional matching to improve matching precision. Both List A->B and List B->A. thus 512 x 512 x 2 (bi-direction) = 524288 to be searched.

Could you please check again XLS below. Note that customer has same concern on (1).

4456.orb estimation.xlsx

Best regards,

Saito

0 Anshu Jain over 7 years ago in reply to Koichi Saito

TI__Guru 56820 points

Hi Saito San,
I will update the excel sheet with updated cycles for matching ( 2.5 cycles). Second point is still under discussion.

Regards,
Anshu

0 Anshu Jain over 7 years ago in reply to Anshu Jain

TI__Guru 56820 points

5611.orb estimation.xlsxHi Saito San,

PFA update performance estimate based on the disucssion we had.

0 Koichi Saito over 7 years ago in reply to Anshu Jain

TI__Intellectual 2525 points

Hi Anshu,

Thanks.
Is it possible to add step-a) to d) and your assumption for feature matching (which you mentioned in email) to XLS as notes by estimator ?

Best regards,
Saito

0 Anshu Jain over 7 years ago in reply to Koichi Saito

TI__Guru 56820 points

Hi Saito San,
As the steps followed might be specific to customer so I will not update here. Can you add this piece of information in the excel which we shared with you. All the information is available in the email chain.

Regards,
Anshu

0 Koichi Saito over 7 years ago in reply to Anshu Jain

TI__Intellectual 2525 points

Hi Anshu,
OK. Thanks for EVE estimation you did so far.

Best regards,
Saito

Processors

Processors forum

RTOS/TDA2P-ACD: ORB performance on EVE : Estimated DDR BW