nimforum mirror - Inference of static[int] in sequences

andrea (orginal) [2015-07-23T13:59:34+02:00] view original

I have a linear algebra library that I am trying to use inside another project. The library heavily makes use of the compiler infering a static[int] parameter representing the dimension of a vector and passing it to BLAS functions, like here

proc `*=`*[N: static[int]](v: var Vector64[N], k: float64) {. inline .} = scal(N, k, v.fp, 1)

Here scal is imported from a C header:

proc scal(N: int, ALPHA: float64, X: ptr float64, INCX: int)
  {. header: header, importc: "cblas_dscal" .}

Now, it seems that the Nim compiler sometimes sees N as a constant and decides to call scal passing just three arguments, so that the compilation fails at the GCC step.

The simplest example that I have been able involves sequences of vectors:

import linalg


proc sum[N: static[int]](vs: seq[Vector64[N]]): Vector64[N] =
  result = zeros(N)
  for v in vs:
    result += v

echo sum(@[
  randomVector(5),
  randomVector(5),
  randomVector(5)
])

This fails with

/home/papillon/esperimenti/example/nimcache/example_example.c: In function ‘HEX2BHEX3D_112413’:
/home/papillon/esperimenti/example/nimcache/example_example.c:205:2: error: incompatible type for argument 2 of ‘cblas_daxpy’
  cblas_daxpy(1.0000000000000000e+00, ((NF*) (w)), ((NI) 1), ((NF*) ((*v))), ((NI) 1));
  ^
In file included from /home/papillon/.bin/intel/composer_xe_2015.0.090/mkl/include/mkl.h:48:0,
               from /home/papillon/esperimenti/example/nimcache/example_example.c:12:
/home/papillon/.bin/intel/composer_xe_2015.0.090/mkl/include/mkl_cblas.h:160:6: note: expected ‘double’ but argument is of type ‘NF *’
 void cblas_daxpy(const MKL_INT N, const double alpha, const double *X,
      ^
/home/papillon/esperimenti/example/nimcache/example_example.c:205:2: error: too few arguments to function ‘cblas_daxpy’
  cblas_daxpy(1.0000000000000000e+00, ((NF*) (w)), ((NI) 1), ((NF*) ((*v))), ((NI) 1));

Any idea how to make this work?

Varriount (orginal) [2015-07-23T23:19:19+02:00] view original

Why do the integers need to be static?

andrea (orginal) [2015-07-23T23:42:59+02:00] view original

Because I want to keep track of vector and matrices dimension at compile time, so that the compiler will refuse to operations where the dimensions do not match.

In fact the compiler does a pretty good job at infering the right dimensions, see the examples in the README and and in the tests. I was hoping to write something that would feel easy to use like numpy, but still avoid common mistakes thanks to the static[int] guarantee

Araq (orginal) [2015-07-24T00:46:42+02:00] view original

Just report the bug properly please. It's not hard to fix. In fact, when I reviewed the code in the compiler dealing with static[T] I was surprised that it actually works, but now I know better. ;-)

andrea (orginal) [2015-07-24T10:06:13+02:00] view original

Thank you, I reported the bug here

I was able to produce an example where the linalg library is not involved.

Mirror of forum.nim-lang.org

1467 :: Inference of static[int] in sequences