5 ms·

Fortran Intrinsics (2018)

39 points by Bostonian 4y ago

inamberclad 4y ago
I know that matrix inversion was deliberately left out of the fortan built in functions
- klodolph 4y ago
  Are there any matrix operations in the built-ins?
  - kergonath 4y ago
    Most numerical operations work on array elements, so things like A+B work as expected with matrices. Not A*B, because that is element-wise multiplication and the matrix product is matmul(). The Euclidean norm norm2() should also work as expected as well.
  - celrod 4y ago
    matmul
- tlb 4y ago
  Probably wise: there are too many ways to do it, each with different behavior for ill-conditioned matrices. GLSL includes an inverse for fixed-sized (2,3,4) matrices, which can return NaNs for ill-conditioned inputs. But it's fine for inverting transformation matrices.
- Extigy 4y ago
  At least there’s always LAPACK if you need it!
spapas82 4y ago
I still remember back in 2004 when I needed to use the bessel j functions for a C++ electromagnetics simulation for the university. Back then, these functions were only available on a 1986 Fortran implementation by DE Amos (https://dl.acm.org/doi/10.1145/7921.214331 https://dl.acm.org/doi/10.1145/7921.214331). I was so delighted when I was able to compile the fortran sources as .obj files and then call them from my C++ program to receive the correct results!
- tlb 4y ago
  Now they {j0, j1, jn, y0, y1, yn} are POSIX-standard and in C/C++'s libm.
  - adrian_b 4y ago
    The standard C Bessel functions accept only real arguments. The algorithm linked by OP is for the Bessel functions extended to complex arguments, which do not exist in the standard C/C++ library.
agumonkey 4y ago
Interesting that the website advertise fortran as natively parallel, was it the case in the 60s too or just some marketing trick to get people to pay attention to this old friend ? (honest question)
- Bostonian 4y ago
  The "natively parallel" claim is made because of features such as "do concurrent" and coarrays, which were added in the Fortran 2008 standard.
pklausler 4y ago
Beware of the examples in the Fortran standard for intrinsic functions like TINY() -- the example results are within the framework of a 32-bit floating-point model that looks an awful lot like IEEE-754 finite numbers (24-bit precision including an implicit normalization bit, 8-bit exponent biased by 127), except for an interpretation that is off by a factor of two. That's right. If you take any of those example calls from the Fortran standard and compile them on any modern processor, your results will differ by a factor of two one way or another. Does the standard warn you about this? Nowhere. Does the standard committee know about this? Yes. Will they fix their documentary floating-point model to conform with the one truly ubiquitous 32-bit floating-point representation left to us? Of course not.

39 points by Bostonian 4y ago

inamberclad 4y ago

I know that matrix inversion was deliberately left out of the fortan built in functions

klodolph 4y ago
Are there any matrix operations in the built-ins?
- kergonath 4y ago
  Most numerical operations work on array elements, so things like A+B work as expected with matrices. Not A*B, because that is element-wise multiplication and the matrix product is matmul(). The Euclidean norm norm2() should also work as expected as well.
- celrod 4y ago
  matmul
tlb 4y ago
Probably wise: there are too many ways to do it, each with different behavior for ill-conditioned matrices. GLSL includes an inverse for fixed-sized (2,3,4) matrices, which can return NaNs for ill-conditioned inputs. But it's fine for inverting transformation matrices.
Extigy 4y ago
At least there’s always LAPACK if you need it!

spapas82 4y ago

I still remember back in 2004 when I needed to use the bessel j functions for a C++ electromagnetics simulation for the university. Back then, these functions were only available on a 1986 Fortran implementation by DE Amos (https://dl.acm.org/doi/10.1145/7921.214331 https://dl.acm.org/doi/10.1145/7921.214331). I was so delighted when I was able to compile the fortran sources as .obj files and then call them from my C++ program to receive the correct results!

tlb 4y ago
Now they {j0, j1, jn, y0, y1, yn} are POSIX-standard and in C/C++'s libm.
- adrian_b 4y ago
  The standard C Bessel functions accept only real arguments. The algorithm linked by OP is for the Bessel functions extended to complex arguments, which do not exist in the standard C/C++ library.

agumonkey 4y ago

Interesting that the website advertise fortran as natively parallel, was it the case in the 60s too or just some marketing trick to get people to pay attention to this old friend ? (honest question)

Bostonian 4y ago
The "natively parallel" claim is made because of features such as "do concurrent" and coarrays, which were added in the Fortran 2008 standard.

pklausler 4y ago

Beware of the examples in the Fortran standard for intrinsic functions like TINY() -- the example results are within the framework of a 32-bit floating-point model that looks an awful lot like IEEE-754 finite numbers (24-bit precision including an implicit normalization bit, 8-bit exponent biased by 127), except for an interpretation that is off by a factor of two. That's right. If you take any of those example calls from the Fortran standard and compile them on any modern processor, your results will differ by a factor of two one way or another. Does the standard warn you about this? Nowhere. Does the standard committee know about this? Yes. Will they fix their documentary floating-point model to conform with the one truly ubiquitous 32-bit floating-point representation left to us? Of course not.