about the number of multipliers that is mentioned in the synthesis report

[kannan2590]

According to this post the question is for the filter written in vhdl code the multipliers that i have used in the code is 164 multipliers .As per the fpga there are total 192 multipliers But the synthesis report it shows 148 multipliers and the filter coefficients are symmmetrical.I doubt whether during the optimization of the code the it has reduced the no of mutipliers.Will it effect the final result in the hardware.I use virtex 4 fpga and this hardware HAS 192 multipliers totally.The simulation results of the filter code shows correct result.What is your opinion on this reduction in the no of multipliers?

 

[TrickyDicky]

if its symmetrical, then you can just pre-add the input values.

 

[ads_ee]

Maybe you're seeing some resource sharing going on. Try turning of any type of resource sharing the the tools and see if the number of DSP blocks used goes back to 164.

 

[kannan2590]

Maybe you're seeing some resource sharing going on. Try turning of any type of resource sharing the the tools and see if the number of DSP blocks used goes back to 164.

how to turn of the resource sharing in the tools.How to find out where the resource sharing option.is it the option and where it can be found out?

 

[std_match]

I doubt that the tools are smart enough to merge multipliers with the same coefficient. Maybe some coefficients are very simple, like a power of 2.
A multiplier is not needed for them.

 

[kannan2590]

I doubt that the tools are smart enough to merge multipliers with the same coefficient. Maybe some coefficients are very simple, like a power of 2.
A multiplier is not needed for them.

I could not understand how a number power of two a multiplier is not needed and could you tell me how to avoid resource sharing among the coefficients .If the coefficients are arranged in LUT will this resource sharing problem can be avoided

 

[permute]

for example:
01001001 = 0x49
0x49*0 = 0 -- no multiplier needed
0x49*1 = 0x49 -- no multiplier needed
0x49*2 = 10010010 =0x92 -- just a left shift of the input, no multiplier needed.

long filters often have a handful of 0-valued coefficients after quantization.

 

[sureshaa]

Hi Kannan,

In your co-efficents, do you have 1 and -1? It is so how many co-eff you have? May be because of that, it may reduce.(For 1 and -1 it doesnt require to multiply right?)

 

[kannan2590]

Hi Kannan,

In your co-efficents, do you have 1 and -1? It is so how many co-eff you have? May be because of that, it may reduce.(For 1 and -1 it doesnt require to multiply right?)

Actually i am having coefficients like FFFE (1111111111111110)and the data multipliying with this coefficints are 90 and -90 only .Whether this reduction in the multiplier will affect the hardware output result

 

[sureshaa]

Hi Kannan,

FFFE (1111111111111110) is nothing but -2.Like permute said, it doesnt required multiplier for that.

 

[kannan2590]

Hi Kannan,

FFFE (1111111111111110) is nothing but -2.Like permute said, it doesnt required multiplier for that.

Actually the question are these xilinx synthesis tools intelligent enough to know these coefficients and optimize the no of multipliers?

 

[sureshaa]

Yeah.. It will try to reduce the multiplier count. If you want to check the tool... just write one eighteen bit multiplier and multiply with "111111111111111110". Now the hardware wont be multiplier. Then change the multiplicant with some random number( which is not 1,-1,2,-2,0, and multiples of 2).Then see the resource utilization.

 

[kannan2590]

Yeah.. It will try to reduce the multiplier count. If you want to check the tool... just write one eighteen bit multiplier and multiply with "111111111111111110". Now the hardware wont be multiplier. Then change the multiplicant with some random number( which is not 1,-1,2,-2,0, and multiples of 2).Then see the resource utilization.

Actually when i kept the numbers like 24,34,36 and other evenn numbers other than the power of two like(2,4,8,16,32.....)it is showing the multipliers used .But when i used(2,4,8,16,32.....) it shows no multipliers are used.

 

[sureshaa]

Actually when i kept the numbers like 24,34,36 and other evenn numbers other than the power of two like(2,4,8,16,32.....)it is showing the multipliers used .But when i used(2,4,8,16,32.....) it shows no multipliers are used.

Yes Dude.. You are correct. I will correct my words. It is not multiples of two, It is two^n.;-)

 

[kannan2590]

Yes Dude.. You are correct. I will correct my words. It is not multiples of two, It is two^n.;-)

Again one more doubt i am having is that if i kept numbers like 7,7,7,7 the no. of multipliers shown is zero.Whether this will have effect on the final output?If yes how to correct it?

 

[std_match]

Multiplication with a power of 2 (1,2,4,8,16, etc) is just a shift. No hardware resources are consumed.

If the coefficient is a power of 2 +/- another power of 2, the multiplication can be implemented with an adder/subtractor. 7 = 2^3 - 2^0

I don't know which tools can take advantage of this.

 

[kannan2590]

Multiplication with a power of 2 (1,2,4,8,16, etc) is just a shift. No hardware resources are consumed.

If the coefficient is a power of 2 +/- another power of 2, the multiplication can be implemented with an adder/subtractor. 7 = 2^3 - 2^0

I don't know which tools can take advantage of this.

But the synthesis tool or any other tool how can it know that it can be written as 7 = 2^3 - 2^0.Whether you have tried it with vhdl code if it is so can you send that code to me.

 

[sureshaa]

Kannan2590,

std_math said it can be.. but he is not sure about that. I think u can do these experiments with tool. It wont take that much time.

 

[std_match]

One interesting special case it to multiply by 10.
Since 10 = 2^3 + 2^1, only one adder is needed.