In an Equation RHS side has values where as LHS side variable is getting assigned to NaN

[acobulareddy]

I have the following equation

TCP(L) = ((A(J)*PLOG + A(J+1))*PLOG + A(J+2))*PLOG + A(J+3) + S41

&+ PLOG*(S3 + PLOG*(S2 + PLOG*S1))

while debugging I came to know

L1INTEGER(4)
J41INTEGER(4)
A(J)281.844320000000REAL(8)
PLOG7.76195750680761REAL(8)
A(J+1)-4428.76670000000REAL(8)
A(J+2)8563.47850000000REAL(8)
A(J+3)76694.1260000000REAL(8)
S4-73755.8052496628REAL(8)
S3-9301.57554486114REAL(8)
S24506.18740882533REAL(8)
S1-283.717539158391REAL(8)

where as TCP(L) is NaN

DIMENSION A(120), AT1(20), AT2(20), AT3(20)
DIMENSION ACP1(20), ACP2(20), ACP3(20), TCP(2)

is saying the array sizes for TCP and A are with in the limit.

when I have set the fortran compiler settings of Data -> Local Variable to All Variable Save it happend not to get that error for a while and again it re appeard after some more calculations.

Please suggest what else are the setting I need to modify to get this work well.

Thanks,
Chandra.

[onkelhotte]

When I run the latter code, I get tcp(1) = 997.687999976726

The equation shows no sign of getting a NaN result. For a better understanding of real number computations, read the three "The Perils of Real Numbers" articels in the Doctor Fortran thread by Steve Lionel: http://software.intel.com/en-us/forums/showthread.php?t=41911

Your hint, that with save the variables you get the NaN more less shows, that some of your variables are not initialized every time you compute tcp.

[cpp]integer(kind=4) l,j
real(kind=8) a(4),tcp(1),s1,s2,s3,s4,plog

l=1
j=1

A(J) = 281.844320000000D0
PLOG = 7.76195750680761D0
A(J+1) = -4428.76670000000D0
A(J+2) = 8563.47850000000D0
A(J+3) = 76694.1260000000D0
S4 = -73755.8052496628D0
S3 = -9301.57554486114D0
S2 = 4506.18740882533D0
S1 = -283.717539158391D0

TCP(l) = ((A(J)*PLOG + A(J+1))*PLOG + A(J+2))*PLOG + A(J+3) + S4 + PLOG*(S3 + PLOG*(S2 + PLOG*S1))[/cpp]

[acobulareddy]

Even I know that there is nothing which leads to resulting the computation to NaN where as still it is happening. I have seen the values of variables and all are having real numbers. I have already posted those values in my question. So there is no question of not declaring some variable.

This equation is coming in my computation for more than 1000 times but at one perticular point this is returning NaN and my computation is not going forward that point onwards. This is straight forward equation and I dont know why this is resulting NaN. I assume memory issues.

I would like to know are there any memory settings like use somuch of memory /increase the memory size etc... I will try by increasing that size and test.

Thanks

Chandra.

[TimP]

I suppose you could consider x87 stack overflow as a sort of out of memory condition, so you should try running with fp stack checking, particularly if you want to compile without SSE2 code generation. I don't get a clear impression what you are thinking of. Other than that, NaN will be produced if one of the operands is NaN, or possibly by some combination with infinity sub-expressions.

The evaluation ought to be somewhat more reliable (and faster), if the polynomial evaluations are merged

TCP(L) = (((A(J)+S1)*PLOG + A(J+1)+S2)*PLOG + A(J+2)+S3)*PLOG + A(J+3) + S41

Setting /fp:precise might encourage the compiler to evaluate the expression in C standard order. The original expression is far too complicated to guess what order of evaluation a compiler will choose when given free rein.

Speaking for myself, I find it more readable if polynomials are always written low order term first. Reversing the order entirely, along with /fp:precise, could give you more assurance that you know the actual order of evaluation without reading asm code.

In case you don't optimize your source code, /pc80 switch might allow more use of extended range with x87 code.l