Question about best performance using ipp CrossCorrNorm function.

[DoS]

Hi,

I have question about function which is documented well here.

1. Is it true that using different RAM types affects the computing speed of this function?

For example.

Using Intel i9-12900K

DDR4 64GB 3200 MHz CL18   vs  DDR5 64GB 4800MHz CL38

Looking at the example above, will the computing speed be faster if I use ddr5 memory?

2. Which is the best Intel processor that will have fastest computing speed per thread?

[ShanmukhS_Intel]

Hi,

Thank you for posting on Intel Communities.

We are working on your issue internally. We will get back to you soon with an update.

Best Regards,

Shanmukh.SS

[Abhinav_S_Intel]

We can better answer your questions if you can provide us with some more information on the input image size, input template size, data type and which function switches are used which means algorithm (auto, FFT, direct) along with correlation used is valid or full. Please post as much information you can provide with the query.

[DoS]

Hi Abhinav_S,

thanks for your answer.

I use:

function = ippsCrossCorrNorm_32f

method = ippAlgFFT

mode =  ippiROIValid

normmode = ippiNorm

data type = float32

values for input image and template = random from 0 to 20

About template, using Python it can be created like:

```

ref_template =  [np.asarray(np.random.randint(5000, 50000)) for _ in range(50000)]

```

I attached txt file with templates created.

You can load it and create values of each template this way:

```

with open(txt_file) as f:

    lines = f.read().splitlines()

ref_templ =  [np.random.randint(low=0, high=20, size=int(ref_size)).astype(np.float32) for ref_size in lines]

```

input image:

```

input_image = np.random.randint(low=0, high=20, size=(500000)).astype(np.float32)

```

[Abhinav_S_Intel]

Analysis needs to be done on how many threads are effective for a particular set of {image size, template size, data type}. The functionality is not threaded in the main IPP library but it has a threaded implementation available with the TL libs. The no. of threads depends on ratios images.width/tpl.width & image.height/tpl.height. The higher is the ratio, the more threads can be used. We can't answer on exactly how many threads can be used.

The memory type (memory speed) may depend on the load size (input images) which if doesn't fit into the 0.5 of LLC (last level cache). There will be a lot of memory-to-memory operations,  therefore the higher is the speed of memory relative to CPU frequency - the more execution speed.