Digital Signal Processing - IP and Algorithms

[argento]

Hello Folks,

I'm running the C Modelexample provided with the package.

I'm using my data. I used Scheduled Scaling, but when I'm going to test Block Fixed-Point I get an error:

ERROR: blk_exp is incorrect: expected 14, actual 7

Some outputs from simulation are incorrect

Output of FFT is unusable, and very far from results I obtain with Scheduled Scaling.

Why is this happening?

How can I solve it?

[ywu]

The blk_exp checking in the C model example is only valid for the input data used in the original example. If you use your own data, the blk_exp will be different, so you will need to remove the blk_exp checking in your C model.

The core automaically adjusts scaling in block floating mode, so blk_exp output can be different for each frame depending on the input data. You will need to left shift the FFT output by blk_exp to reconstruct the output data.

---

argento wrote:

Hello Folks,

I'm running the C Modelexample provided with the package.

I'm using my data. I used Scheduled Scaling, but when I'm going to test Block Fixed-Point I get an error:

ERROR: blk_exp is incorrect: expected 14, actual 7

Some outputs from simulation are incorrect

Output of FFT is unusable, and very far from results I obtain with Scheduled Scaling.

Why is this happening?

How can I solve it?

---

[argento]

Ok, I'll remove that check.

But anyway, I'd be glad if you can teach me why blk_exp should be equal to nfft, if it is possible.

Thanks in advance

[argento]

Ok, I'll remove that check.

But anyway, I'd be glad if you can teach me why blk_exp should be equal to nfft, if it is possible.

Thanks in advance

[ywu]

The blk_exp doesn't always equal to nfft. It happens to be nfft in the provided example when the input is a constant 0.5.

---

argento wrote:

Ok, I'll remove that check.

But anyway, I'd be glad if you can teach me why blk_exp should be equal to nfft, if it is possible.

Thanks in advance

---

[ywu]

Ok, I'll remove that check.

But anyway, I'd be glad if you can teach me why blk_exp should be equal to nfft, if it is possible.

Thanks in advance

[argento]

Ok, I managed to use BFP.

One problem was I was using bad data.

Block Floating-Point seems to work now.

What is faster for the LogiCORETM IP Fast Fourier Transform v7.1,

to use a Scaling Schedule or BPF ?

[ywu]

If by "faster" you mean Fmax, the difference between scaled and block floating point is negligible. You can easily generate two netlists and compare the Fmax yourself.

---

argento wrote:

Ok, I managed to use BFP.

One problem was I was using bad data.

Block Floating-Point seems to work now.

What is faster for the LogiCORETM IP Fast Fourier Transform v7.1,

to use a Scaling Schedule or BPF ?

---