TMS320F280025: Fixed-point division by FINTDIV vs floating-point division by TMU

Wayne Huang

Genius 16325 points

Part Number: TMS320F280025
Other Parts Discussed in Thread: C2000WARE

Dear Champs,

I am asking this for our customer.

The user used to use fixed-point code and want to migrate the fixed-point code onto our C2000 MCU.

We are aware that floating-point division is done by TMU and fixed-point division is done by FINTDIV.

From the app note "Fast Integer Division – A Differentiated Offering From C2000 Product Family"

https://www.ti.com/lit/an/spracn6/spracn6.pdf?ts=1622433172943

It takes 13-16 cycle for a traditional fixed-point division.

For TMU, it takes about 5 cycles for a floating-point division.

It appears floating-point division by TMU is faster than fixed-point division by FINTDIV.

Therefore,

The user wonders which one of below is faster for division?

Do you have any suggestion or preference or comment for the user when they consider implementation?

1) By fixed-point division and using FINTDIV only

Int a, b, c;

c = a/b;

2) By converting to floating-point division and using TMU.

Int a, b, c;

float a1, b1, c1;

a1 = (float)a;

b1 = (float)b;

c1 = a1/b1;

c = (int)(c1*scaling_value);

Wayne Huang

over 4 years ago

0 Shanty over 4 years ago

TI__Expert 5410 points

Hi Wayne,

Floating point division using the FASTRTS library using TMU takes 25 cycles for single precision. This is not including the overhead to convert to float, which takes and additional 3 cycles.

For your usecase, I would suggest using the FASTINTDIV library directly.

-Shantanu

0 Wayne Huang over 4 years ago in reply to Shanty

TI__Genius 16325 points

Dear Shantanu

I am confused.
What do you mean by 25 cycles as we think it takes 5 cycles for floating-point division in TMU? Why does it differ a lot?

Would you please explain it?

Wayne

0 Shanty over 4 years ago in reply to Wayne Huang

TI__Expert 5410 points

Wayne,

I am looking at the benchmarking information provided in the FASTRTS docs in c2000ware for single precision division. WHere did you get the 5 cycles information from?

-Shantanu

0 Wayne Huang over 4 years ago in reply to Shanty

TI__Genius 16325 points

Dear Shantanu,

I was referring to the instruction, DIVF32, of Table 8-6. TMU Supported Instructions of the datasheet.

https://www.ti.com/lit/ds/symlink/tms320f280025c.pdf?ts=1622684505894&ref_url=https%253A%252F%252Fwww.ti.com%252Fproduct%252FTMS320F280025C

Is there anything wrong?

Wayne Huang

+1 Shanty over 4 years ago in reply to Wayne Huang

TI__Expert 5410 points

Wayne,

My apologies. The API I had referenced above is from the standard RTS library.

The DIV32 function for float division using TMU and --fp_mode=relaxed takes 5 cycles, the MOV32 operations preceding it take around 3 cycles total and conversion from int to float (I32TOF32) takes around 5 cycles per operation.

For a 32 bit FASTINTDIV, traditional division for 132 / i32 takes around 13-14 cycles for the same.

Hence, for your usecase it is better to use FASTINTDIV to avoid the overhead of conversion to float.

-Shantanu

C2000™︎ microcontrollers

C2000 microcontrollers forum

TMS320F280025: Fixed-point division by FINTDIV vs floating-point division by TMU