Hi, Suppose I have a complex multiplier from the MegaWizard with inputs having 16 bits each and the outputs also having 16 bits. Because 16*16 equals 32 bits, would restricting the output to 16 bits output the Most Significant Bits (MSBs), i.e. would it output [31 downto 16] bits of the total 32 bits? Also, would the above reasoning be the same for the altmult_add MegaWizard block too? Is it common for all the Quartus arithmetic in general (to output the MSBs)? If not, how can I force only the MSBs to be outputted? Thank you.

The arithmetic functions should discard LSBs if it allows you less than maximum width. If in doubt, you may keep it maximum width then choose by hand the bits you want to keep and discard other bits.

Alright, I will go with the second option....Thank you.

Suppose the 32-bit output is "00000011010100100000111100001010"...how would you ignore the first 6 0-bits and then output the 16 MSB "1101010010000011"?

When discarding most significant bits, you should use saturation logic to avoid overflow. For unsigned multiply, if any of the 6 bits on the left is set, the output should be forced to maximum value. For signed multiply, to min- or maxint, depending on the sign.

Are only MSBs outputted if bits are restricted?

11 Replies

Altera_Forum
Honored Contributor
15 years ago
The arithmetic functions should discard LSBs if it allows you less than maximum width. If in doubt, you may keep it maximum width then choose by hand the bits you want to keep and discard other bits.
Altera_Forum
Honored Contributor
15 years ago
Alright, I will go with the second option....Thank you.
Altera_Forum
Honored Contributor
15 years ago
lpm_mult does discard the LSB.
Altera_Forum
Honored Contributor
15 years ago
Suppose the 32-bit output is "00000011010100100000111100001010"...how would you ignore the first 6 0-bits and then output the 16 MSB "1101010010000011"?
Altera_Forum
Honored Contributor
15 years ago
When discarding most significant bits, you should use saturation logic to avoid overflow. For unsigned multiply, if any of the 6 bits on the left is set, the output should be forced to maximum value. For signed multiply, to min- or maxint, depending on the sign.
Altera_Forum
Honored Contributor
15 years ago
Unfortunately, I am relatively new to VHDL and Altera Quartus, so I couldn't really comprehend your statement above...

I was actually coding something along these lines:

BEGIN
PROCESS (clock)
BEGIN
IF (clock = '1') THEN
FOR i IN 31 DOWNTO 0 LOOP
IF datain(i) /= '0' THEN
out_msb <= datain(i DOWNTO (i-15)); -->line 41
END IF;
END LOOP;
END IF;
END PROCESS;
END rtl;

But compiling this gives me an error:
"error (10453): vhdl error at msb_output.vhd(41): right bound (-1) of slice must belong to range (31 downto 0) of corresponding object"

What I was trying to do is loop until I get the first '1' from the left, then take the 16-bits starting from the first '1' as my 16-MSB output.

Help and guidance will be highly appreciated.
Altera_Forum
Honored Contributor
15 years ago
You cannot easily drop a selective number of 0's, you usually just drop a fixed N bits and do the saturation logic as FvM describes, maybe even adding an overflow bit to let other things know that a number has saturated and not actually come out to max. The big problem would be knowing what power of 2 the MSB represents if you tried that, along with all the extra logic required to count the number of 0s and output the next 16.
Altera_Forum
Honored Contributor
15 years ago
--- Quote Start ---
Unfortunately, I am relatively new to VHDL and Altera Quartus, so I couldn't really comprehend your statement above...

I was actually coding something along these lines:

BEGIN
PROCESS (clock)
BEGIN
IF (clock = '1') THEN
FOR i IN 31 DOWNTO 0 LOOP
IF datain(i) /= '0' THEN
out_msb <= datain(i DOWNTO (i-15)); -->line 41
END IF;
END LOOP;
END IF;
END PROCESS;
END rtl;

But compiling this gives me an error:
"error (10453): vhdl error at msb_output.vhd(41): right bound (-1) of slice must belong to range (31 downto 0) of corresponding object"

What I was trying to do is loop until I get the first '1' from the left, then take the 16-bits starting from the first '1' as my 16-MSB output.

Help and guidance will be highly appreciated.
--- Quote End ---

In HDL, for loops are unrolled at synthesis time - loops do not exist on hardware. Whats happening here is that it's getting to i = 14 and then it tries to take the slice 14 downto -1. What the synthesisor is creating is a 32 to 1 mux with all the different inputs (31 downto 16), (30 downto 15) etc available, and then all the bits are put through logic gates to create select lines that select the appropriate 16 bit word. But as you've seen, your code tries to create and input to this MUX with an invalid range.
Altera_Forum
Honored Contributor
15 years ago
Then what would be a better way to code this? Could I stop the loop once I come across the first '1'?

Or would a while-loop be better here? How would I implement it?

Appreciate all your valuable help.
Altera_Forum
Honored Contributor
15 years ago
--- Quote Start ---
Unfortunately, I am relatively new to VHDL and Altera Quartus, so I couldn't really comprehend your statement above...
--- Quote End ---

My answer isn't particularly related to VHDL nor Quartus. It's a general statement about mathematical operations. Instead of trying to explain the matter further, let me say: Get you a pencil and paper (or may be a pocket calculator for convenience) and try to figure it out.

To turn your above code into legal VHDL and also perform the apparently intended operation in all cases, you would change it like this:
FOR i IN 31 DOWNTO 15 LOOP IF datain(i) /= '0' OR i =15 THEN out_msb <= datain(i DOWNTO (i-15)); -->line 41 EXIT; END IF; END LOOP;

It takes not less than 185 Cyclone III logic cells, but gets a result in one cycle. By designing a sequential shift, you can considerably reduce the effort, but have to wait up to 16 cycles.

You'll find a similar logic in a floating pointer normalizer. In addition, you would output the actual shift that's used for the exponent. In fixed point arithmethic, you try to avoid variable scaling, saturation logic, as mentioned above, may be necessary however.

Forum Discussion

Are only MSBs outputted if bits are restricted?

11 Replies

Recent Discussions

Does direct connection of two AXI4 Masters to DDR Slave support auto-arbitration on Agilex 5?

Agilex5 Engineering Silicon

Arria 10: Remote Update Watchdog unpredicted behavior

MAX 10. No 3.0 V Schmitt Trigger I/O standard

CYCLONE 10LP Decouple capacitors