28x FPU: Question on conversion instructions from float to integer

Hi Johannes!

You can refer to this manual 1817.TI_TMS320C28xFloatingPointUnit.pdf (particularly page 52). Also this thread will be useful perhaps

http://e2e.ti.com/support/microcontrollers/c2000/f/171/p/79232/286417.aspx#286417

Regards,

Igor

Hi Igor,

thank you for your response. The PDF you have linked in your response is SPRUEO2A which I already had found.

In the description on page 55 it says

Convert the 32-bit floating-point value in RbH to an unsigned 16-bit integer and round to
the closest even value. The result will be stored in RaH. To instead truncate the
converted value, use the F32TOUI16 instruction.

In the examples there is:

MOVIZ R5H, #0x412C  ; R5H = 0x412C
MOVXI R5H, #0xCCCD  ; R5H = 0xCCCD
                    ; R5H = 10.8 (0x412CCCCD)
F32TOUI16R R6H, R5H ; R6H (15:0) = F32TOUI16round (R5H)
                    ; R6H (31:16) = 0x0000
MOVF32 R7H, #-10.8  ; R7H = -10.8 (0x0xC12CCCCD)
                    ; <-- F32TOUI16R complete,
                    ; R6H (15:0) = 11.0 (0x000B)
                    ; R6H (31:16) = 0.0 (0x0000)

F32TOUI16R R0H, R7H ; R0H (15:0) = F32TOUI16round (R7H)
                    ; R0H (31:16) = 0x0000
NOP                 ; 1 Cycle delay for F32TOUI16R to complete
                    ; <-- F32TOUI16R complete,
                    ; R0H (15:0) = 0.0 (0x0000)
                    ; R0H (31:16) = 0.0 (0x0000)

which essentially states that 10.8 is rounded to 11, which is odd and not even. So my trust in this documentation was undermined. (I assume the even in the description refers to xxx.5 floats.)

In this document I didn't find detailed information on what the mentioned F32TOI16round() function exactly does.

From the second part of the example I figure that -10.8 is rounded to 0. Does that hold for all negative values?

What about positive values above 65535?

Regards, Johannes

Hi Johannes!

Honestly I don't unerstand your confusion about conversion 10.8 -> 11. What integer value would you like to get at this case?

Regards,

Igor

Hi Igor,

11 is fine, but the description states that the instruction rounds to the nearest even value and 11 is not even.

Regards, Johannes

Hi!

I think the sentence "round to the closest even value" means the so-called banker's rounding : 2,5 → 2, 3,5 → 4, 11.5→ 12, 10.5→ 10 etc.

Regards,

Igor

It should probably be called Round to nearest, tie to even.  The value is rounded to the nearest integer.  If it happens to fall 1/2 way between the two values, then even is selected.  www.ti.com/lit/SPRAAN9 has an example

Hello Lori!

Thus my above interpretation is correct or not? (Round to nearest (tie to even) column at your Table)

Thanks,

Regard,

Igor 

Igor Gorbachev said:

Thus my above interpretation is correct or not?

Yes.

Hi Lori,

let me rephrase my original question: What is the result (RaH(15:0)) of

F32TOUI16R RaH, RbH

when

• the value of RbH is > 65535.5,
• the value of RbH is < -0.5

an where do I find the official documentation on this?

Similar questions arise for the other conversion instructions (with or without rounding).

Regards, Johannes

Hi Johannes!

"0" is an even number. Thus in above both cases you will get 0 because it is nearest even value.

Regards,

Igor

Thank you, Lori. This is the information I was looking for.

Regards, Johannes