TDA3LX: more efficiently lookup action

Hello!

I don't know for sure what other processing you are doing there, but at least subframe extraction should be doable by EDMA alone, without CPU. One may configure completion interrupt, so DSP gets notified once extraction done and comes with actual processing algorithm.

Hi rrlagic, 

I'm not quite sure what you mean,

maybe I describe too much information,

To be simple, if I execute the code below with dsp, it will take 100ms

int idx = 0;
for(i=0; i<imageH; i++)
{
    for(j=0; j<imageW; j++)
    {
        outPutImg[idx] = Lut[idx];
        idx++;
    }
}

I want to shorten the execution time of this program,

How can I improve?

any method like use edma or eve kernel?

SJay

Hello!

In original post you stated DSP is pretty busy doing other jobs.

Then what I suggest is to offload subframe extraction work to EDMA controller. It will not happen faster. In fact, it may even take little longer depending on memory traffic in your system. What is important, this work consumes no DSP cycles, and it can be run in parallel, in background with other jobs.

The example you show looks too much stripped to me. Really, if 'outPutImg' and 'Lut' both are linear arrays, why do you process them in nested loops? So I presume there is some more logic behind the scene and just telling EDMA can handle 2D transfers, like subframe extration/insertion. If that is merely lienar array copy, still EDMA can be used to offload this job from CPU. So whatever time it takes to make actual transfer, it will make no impact on DSP load. You'll need to issue just few configuration writes.

Just for this case there is QDMA unit in each EDMA controller. Configuration of QDMA transfer takes from 8 to 1 dword writes by processor.

I'm stepping on thin ice of guessing now, so if you're not familiar with DMA, let me give you very small example. Suppose we have 2 linear arrays to copy, like

unsigned char src[SIZE];
unsigned char dst[SIZE];

Then you program QDMA controller with following settings

QDMA_OPT = options_bitfiled;
QDMA_SRC = src;
QDMA_AB_CNT = MAKE_A_B_CNT_MACRO(SIZE, 0);
QDMA_DST = dst;
// zero to unused regs

There is some more details on which config word triggers the transfer, but in general moving block of data is that simple.

Hi rrlagic,

Thank you for your comments,

In fact, I have used EDMA, and I also know that it is suitable for large-scale array handling.

I found that my sample code description is a bit wrong,

Should be changed as below,

int idx = 0;
for(i=0; i<imageH; i++)
{
    for(j=0; j<imageW; j++)
    {
        outPutImg[idx] =inPutTexture[ Lut[idx] ];
        idx++;
    }
}

As shown in the code,

When getting a value from the inPutTexture by lookup the value according to the lookup table,

it is not suitable to use EDMA to operate.

Well,

I felt the example was too simple. Yet again, why do you use nested loops over linear array?

Next, what is contents of LUT? Does it make any back and forth moves over input?

Hi rrlagic,

In fact, I will develop an AVM system that does not use GPU on rtos.

so, if it is not designed as a lookup table method,

the AVM system is difficult to achieve real-time,

The lookup table action is used to obtain the texture pixels value corresponding to the model point coordinates.

SJay

Hi rrlagic,

ok, I understand the advice you gave,

I will design and implement according to this idea,

thank!

Curious to ask, are you Ti FAE?

Because your account doesn’t look like it, but it’s the first time I encountered a reply from a non-Ti FAE : O

Hi again,

Glad to be of help.

No, I am not formally affiliated with TI, another developer like you, just had experience with TI processors.

Any further questions on this thread? Let us know if we can close it.

Regards,

Brijesh

Hi,

you could close it,

if there is any further questions, i will create another thread, thx.