Hello Dan,

Can you show me how you set up the arrays which are pointed to in the FFT object (both input and output), as well as how they are aligned in memory?

Best regards,

Omer Amir

Arrays defined as follows:

in ".cmd" file:

FFT_buffer_1 : > RAMGS0123, ALIGN = FFT_ALIGN
FFT_buffer_2 : > RAMGS0123, ALIGN = FFT_ALIGN

where FFT_ALIGN is defined in C2000 Linker->Advanced Options->Command File Preprocessing: 

in ".c" source file:

#pragma DATA_SECTION(cfft_input, "FFT_buffer_1")
#pragma DATA_SECTION(cfft_output, "FFT_buffer_2")

float cfft_input [PS_CFFT_BUFF_SIZE]; // enough to hold [512 complex 32-bit floats]
float cfft_output [PS_CFFT_BUFF_SIZE]; // enough to hold [512 complex 32-bit floats]

where: PS_CFFT_BUFF_SIZE = 1024

in ".map" file:

page address name
---- ------- ----
0 0000c000 cfft_input
0 0000c800 cfft_output

You can see the FFT structure *(hnd_cfft) state above in the highlighted yellow regions.

Anything else you need?

So, text you quoted (above) shows the CurrentOutPtr should point to the result (for odd number of stages), then you say "the assembly note says the opposite"??? Where is this "secret" note because I am looking at other fft routines and I do NOT see this in my library assembly code!

While you are correcting things,  in CFFT_f32t.asm, it says:

;; * MagBuf and PhaseBuf are FFTSize/2 in length
;; * MagBuf and PhaseBuf are not used by this function.
;; They are only used by the magnitude and phase calculation functions.

This seems incorrect, because it (otherwhere) says size of mag and phase should be (FFTSize/2 +1).

Hello Daniel,

The assembly file is called CFFT_f32t.asm (name of the function), and it can be found in the library source file directory: C2000Ware_4_XX_XX_XX\libraries\dsp\FPU\c28\source\fpu32\fft. I believe the note in the example is correct, because looking at the assembly the pointers do get swapped after each stage. Therefore, an even number of stages will make CurrentInPtr point back to the input.

Daniel Staver said:

So, text you quoted (above) shows the CurrentOutPtr should point to the result (for odd number of stages)

This is not quite right, CurrentInPtr should always point to the final output, but the array it's pointing to changes depending on the number of stages in the FFT (so CurrentInPtr can be pointing to either the input or the output, but it will always be pointing to the final output).

Daniel Staver said:

This seems incorrect, because it (otherwhere) says size of mag and phase should be (FFTSize/2 +1).

Can you tell me where you found where it says FFTSize/2+1? I want to double check some things.

Best regards,

Omer Amir

OK, so I am completely confused by the seemingly contradictory comments. I was just trying to follow the guides.

"FPU DSP Software Library User's Guide, FPU_SW_LIB_UG.pdf" Sec-5.4.2.18 says:

CurrentInPtr and CurrentOutPtr to keep track of the switching. The user can
determine the address of the final output by looking at the CurrentOutPtr.

The canned example fft projects always seem to check "CurrentOutPtr at the end of an FFT.

In example project "cfft_f32", file "dsp_cfft.c", lines 168-171, it says:

// If number of Stages is ODD,
// currentInPtr=test_input, currentOutPtr=test_output
// If number of Stages is EVEN,
// currentInPtr=test_output, currentOutPtr=test_input

Buried deep in the assembly code is your secret comment on line-158:

; Final output saved in memory pointed by CurrentInPtr

So, is that ALWAYS the case and the other comments are wrong?

As for the (FFTSize/2 + 1) issue, I thought because the unpack function seems to make a result of (N/2+1), I thought that applied to mag and phase, too, no?

See project "rfft_f32", file: "dsp_rfft.c", lines 116-122:

// Run the unpack function on the N/2 point complex spectrum
// to get the N/2+1 point complex spectrum of the N point real
// data. The unpack function will generate the complex
// spectrum from F(0) to F(N/2).
// The pointer, p_currOutput points to the complex spectrum
// of the real N-point input data.
CFFT_f32_unpack(hnd_cfft);

lines: 127-129

// The golden vector has the full spectrum of the real data, i.e. F(0)
// to F(N-1) but the unpack routine however only calculates the
// spectrum from F(0) to F(N/2)

Hi Daniel,

Daniel Staver said:

"FPU DSP Software Library User's Guide, FPU_SW_LIB_UG.pdf" Sec-5.4.2.18 says:

I confirmed this in the guide, but after looking at the cfft_f32t.asm file comments specifically within the functions (which is line 158 as you mentioned), I believe this may be a mistake (I will file the appropriate change).

Daniel Staver said:

In example project "cfft_f32", file "dsp_cfft.c", lines 168-171, it says:

If you read the comment prior to this within the same block, you'll see that this set of comments is specifically referring to the CFFT_f32_mag function (for calculating magnitude), which operates differently than the FFT function.

Daniel Staver said:

See project "rfft_f32", file: "dsp_rfft.c", lines 116-122:

This specific example and this function are used for converting complex FFT output into a real FFT output. Per the user guide, "Unpack the N-point complex FFT output to get the FFT of a 2N point real sequence." The cfft_f32t example is probably a more apt example to go by if you aren't looking to convert the FFT from complex to real values, but let me know if this is what you're trying to implement.

Best regards,

Omer Amir