how to DSP-code "signed square root"?

[steph_tsf]

if x positive
then x = sqrt(x)
else x = - (sqrt (-x))
end

I've just completed a 64-tap Widrow-Hoff LMS Machinery. It is written in DSP (not in Green, not in Ruby, not in x86 Assembly). The "textbook" variant is doing fine. I would like to experiment a "signed square root" variant, that's "softening" the error signal using the above mentioned "signed square root" function, and also "softening" the incoming signal, before one is combining them using a multiplication.

Code: Select all

mu2err = mu * 2.0 * err;
a00=leak*(a00+(mu2err * m00));
a01=leak*(a01+(mu2err * m01));
a02=leak*(a02+(mu2err * m02));
a03=leak*(a03+(mu2err * m03));
a04=leak*(a04+(mu2err * m04));


Thus, I need to compute :
a00=leak*(a00+SIGNEDSQUAREROOT(mu2err * m00));
a01=leak*(a01+SIGNEDSQUAREROOT(mu2err * m01));
etc.

intelliger 64 v1.0 leaky.fsm

(2.11 MiB) Downloaded 1241 times

intelliger 64 v1.0 leaky.jpg (650 pix).jpg (74.96 KiB) Viewed 17403 times

Any help much appreciated
Have a nice day

[pshannon]

Just taking a jab here. Did you actually check the results of X? When you expect a negative signed result, are you actually getting this? Sometimes if then and variable assignments have given me incorrect results and I created a work around for testing to see if the added steps resolve the issue even though the logic appears 100% correct. I am only assuming your equation is 100% correct and should work. The other Assumption: if x positive true/false works? I did not try your attached .fsm and if I am way off base, my bad. I am only trying to give you ideas.
Below will force signed x.

Try x=sqrt(x*-1.0) * -1.0
and/or
if x<0 then
x = - (sqrt (-x))
else
x = sqrt(x)
end

[martinvicanek]

In code:

Code: Select all

streamin x;
streamout y;

float abs, sgn        // bitmasks

stage(0){
  abs = 3.2|1.4;      //  abs bitmask 01111111 11111111 11111111 11111111
  sgn = -0;           // sign bitmask 10000000 00000000 00000000 00000000
  }

stage(2){
  y = (sgn&x)|sqrt(abs&x);
  }

ASM:

Code: Select all

streamin x;
streamout y;

int abs=2147483647;  //  abs bitmask 01111111 11111111 11111111 11111111
int sgn=-2147483648; // sign bitmask 10000000 00000000 00000000 00000000

movaps xmm0,abs;
andps xmm0,x;
sqrtps xmm0,xmm0;
movaps xmm1,sgn;
andps xmm1,x;
orps xmm1,xmm0;
movaps y,xmm1;



[steph_tsf]

Try x=sqrt(x*-1.0) * -1.0

Unfortunately, I don't understand your suggestion. In Flowstone DSP Code Component, conditional statements must be implemented using a mask.
For example, the following Flowstone DSP Code: x = x - (x >= 1) & 1.0;
is equivalent to the following C/C++ code: if( x >= 1 ) x = x – 1.0;

The following cumbersome Flowstone DSP code is doing what I am requiring:

Code: Select all

streamin x;
streamout y;
y=sqrt(abs(x))*(((x<0)&-1.0)+(x>0)&1.0);


The corresponding assembly code is:

Code: Select all

streamin x;streamout y;float smIntVarTemp=0.0;
float smIntVarZero=0;
float F0=0;
float FM1=-1;
float F1=1;
movaps xmm0,x;
movaps smIntVarTemp,xmm0;
cmpps xmm0,smIntVarZero,6;
andps xmm0,smIntVarTemp;
addps xmm0,xmm0;
subps xmm0,smIntVarTemp;
sqrtps xmm0,xmm0;
movaps xmm1,x;
cmpps xmm1,F0,1;
movaps xmm2,xmm1;
andps xmm2,FM1;
movaps xmm3,x;
cmpps xmm3,F0,6;
movaps xmm4,xmm3;
andps xmm4,F1;
addps xmm2,xmm4;
mulps xmm0,xmm2;
//Assignment> sLeft=xmm0
movaps y,xmm0;


Looks like a joke
Is there a more efficient approach, still in Flowstone DSP Code Component?
Any help much appreciated

[steph_tsf]

Martin, I didn't notice your kind reply dated Wed Mar 18, 2020 9:40 pm.
Looks brilliant.

Code: Select all

a00=leak*(a00+(mu2err * m00));
a01=leak*(a01+(mu2err * m01));
a02=leak*(a02+(mu2err * m02));
etc.


can thus become in stage (2):

Code: Select all

x=mu2err*m00;
a00=leak*(a00+((sgn&x)|sqrt(abs&x)));
x=mu2err*m01;
a01=leak*(a01+((sgn&x)|sqrt(abs&x)));
x=mu2err*m02;
a02=leak*(a02+((sgn&x)|sqrt(abs&x)));
etc.


Is it possible to also pipe-line the various mu2err*m00, mu2err*m01, etc. calculations, for regaining a "single line" layout?

[pshannon]

When I responded it was to quickly and I thought it was a logic issue you were having or the signed x was not working correctly. I was leaning towards a bug in FS at first. I still know C & ASM, but I have not worked with the DSP code side yet. Sorry for the confusion.

[steph_tsf]

I have not worked with the DSP code side yet.

Me to. This is the first time I am digging into Flowstone DSP Code Component, deeper than computing y = a * x + b.

Attached, is a working implementation of Martin Vicanek solution:

signed square root.fsm

(8.18 KiB) Downloaded 1261 times

http://www.dsprobotics.com/Files/V3/User%20Guide.pdf chapter 9 page 240 doesn't list:
- the "vertical bar" operator,
- the "abs" (absolute value) function,
- the "square root" (sqrt) function.
Question: Is there a supported operators list somewhere?
Question: Is there a supported functions list somewhere?

Upon reading page 242, I realized that one can declare and access indexed arrays.
Question: is this really supported, and reliable, because on Flowstone 3.0.4, this appears to be not supported.
Question: can somebody please publish a circular buffer management example, whose length is 64?

Upon reading page 243, I realized that memin() creates a Mem type connector on the component so that one can pass in data from outside.
Question: is this really supported, and reliable, because on Flowstone 3.0.4, this appears to be not supported.
Question: can somebody please publish a memin() example?
Question: is there a memout()?

Upon reading page 243, I realized one can program loops.
Question: Is it allowed to rely on a loop counter, that's a variable instead of a constant?

Considering the lack of "normal" conditional statements in DSP coding, what Google search keyword do I need to specify for locating the utility that's converting a "if (condition) then (true branch) else (false branch) end" code sequence into a DSP code sequence?

Any help, much appreciated
Have a nice day