tms320c5505: hwafft output

hello..plz solve my issue..why i am getting such output.. i have followed your  errata too...

To execute the HWAFFT routines from the ROM of the DSP, the programmer must satisfy memory allocation restrictions for the data and scratch buffers. See the device-specific errata for an explanation of the restrictions and workarounds:

• TMS320VC5505/VC5504 Fixed-Point DSP Silicon Errata (Silicon Revision 1.4) [literature number SPRZ281]

• TMS320C5505/C5504 Fixed-Point DSP Silicon Errata (Silicon Revision 2.0) [literature number SPRZ310]

• TMS320C5515/C5514 Fixed-Point DSP Silicon Errata (Silicon Revision 2.0) [literature number SPRZ308

still i am getting the same issue..please solve this.

plzzzz somebody reply me...

Hi Kanchan,

It looks like your numbers are in a different fixed-point notation than what is expected by the HWAFFT accelerator.

The HWAFFT inputs and outputs 2's complement 16-bit fixed-point S16Q15 fractional numbers ranging from -1 to 0.9999.

The most significant bit is the sign bit followed immediately by the decimal point and 15 fractional bits (S.QQQ QQQQ QQQQ QQQQ)

Examples of converting from S16Q15 hex to decimal:
Hex               Decimal
0x8000    = -1.000000000
0x8001    = -0.999969482
0xC000   = -0.500000000
0x0000   =   0.000000000
0x0001   =   0.000030518
0x4000   =   0.500000000
0x7FFF  =   0.999969482

I dont understand the numbers that you are getting as output (like -12 and 65530). These must be signed 32-bit numbers. A signed 16-bit (without any decimal point) has a range from -32768 to 37767.

Anyway, you should interpret the output as S16Q15 fractional numbers ranging from decimal -1 to 0.9999.

Given that your inputs {1,2,4,8,16,32,64,128,} are actually {0.000030518, 0.000061035, 0.000152588, 0.000274658, 0.000518799, 0.001007080, 0.001983643, 0.003936768} (see below conversion table), does the FFT result now make sense? You will have to interpret the numbers again in the S16Q15 format.

Maybe you can post the hex numbers in 16-bit format.

You can use matlab to convert to fixed point numbers with code below:

Nsig = 1024;
fs = 16000;
sig = zeros(1,Nsig);
n = 0:(Nsig-1);    % discrete-time grid 
%Add a pure sinusoidal tone
fk = 1000;
sig = .5 * sin(2*pi*n*fk/fs);        %sinusoid

sig_fixed = fi(sig,1,16,15);
sig_fixed_hex = sig_fixed.hex;

I've attached the excel sheet that created the table below:

/cfs-file/__key/communityserver-discussions-components-files/791/6574.S16Q15-fixed-point-numbers.xlsx

Hope this helps,
Mark

Hello Mark,

Thank u so much for the most awaited response from ur side...

thanks for the valuable information.

i am getting the desired output i mean theorotical output (but the rounded integers)if i set the scale flag as 1 i.e. scale_flag = NOSCALE_FLAG;

but now the problem is , HWAFFT has inbuilt function as 

Uint16 hwafft_8pts (
Int32 *data,
Int32 *scratch,
Int32 *duplicate_scratch,
Int32 *duplicate_data,
Uint16 fft_flag,
Uint16 scale_flag
);

where  i can pass only the integer value..what if i want both input and output as float..

if i forcefully do it i couldnt seperate imaginary and real part..if i do so using typecasting the output is completely different..

plz suggest me what do i do?

Hi,

You are welcome. Sorry for the delay.

The HWAFFT accelerator cannot work with float. It only uses 16-bit fixed point ranging from -1 to 0.9999.

You could convert from float to 16-bit int, convert, and then back to float. But the C55xx CPU is not optimized for floating point and it will require several more cycles to perform floating point operations (in software).

I recommend you stick with fixed point with this processor.

Hope this helps,
Mark

Thanks Mark...!

u mean i have to use typecasting only to convert int to float..still i will not get the exact result..then in this case i guess i need to change the processor. :(

thank u Mark for the suggestions!

my application is test and measurement instrument application where i need to calculate phase angle and frequency of the mixed signals..so what do i do..do i go with the same processor..i have already bought tms320c5505 evm usb stick..to develop the application. and really..i will be very difficult to change the processor but i dont get the desired output i have to do so..right? or is there any other option?

DSP processing in C language using CCV7.4..results get matches now but if theoretical result is 57.98 then it gives 57

and my application is to measure tan delta where i have to measure the phase angle and its precision should be of 0.00001 

so if fft results will be like this i will not able to achieve the desired results.

are sure by using DSPLIB i can achieve float or exact result.