int64 version of openBLAS

[access2rohit]

Hi,
I am trying to build MXNet with int64 version of openBLAS. I wanted to know that apart from INTERFACE64=1 and BINARY=64 flags. Is there a binary version of openBLAS available that supports int64 ?
So I can install it like sudo apt install libopenblas64-dev sth like this ?

[martin-frbg]

As far as I know there is a libopenblas64-dev package provided by debian-science, though it may not yet be at the very latest (0.3.10) release

[access2rohit]

@martin-frbg Does it support int64_t API signatures ? Or is just for 64-bit architecture platforms?

[martin-frbg]

I doubt they would add a "64" suffix just for the architecture, but as I said I am not directly involved in providing that package.

[brada4]

It is rarity, you might need to do similar tricks outside julia:
apache/mxnet#16125 (comment)

[access2rohit]

@brada4
https://packages.debian.org/sid/metapackages/libopenblas64-dev

Here it says: "This metapackage provides a 64-bit indexing version" which means this is good for inputs with sizes > (2^31-1)

[brada4]

internal dimensions math is not uint64 perfect in int32 build, probably matrix dimensions over int16 (32k pessimistically) would flop inexplicably. ideally single dimension of 2^31-1 should be working if anyone had time to review through all dimension math inside.

[access2rohit]

@brada4 i didn't understand this part "internal dimensions math is not uint64 perfect in int32 build, probably matrix dimensions over int16 (32k pessimistically) would flop inexplicably"

[martin-frbg]

Seems he was still referring to what happens when one does not use the INTERFACE64=1 version for large problems.

[access2rohit]

@martin-frbg quick question: If i build openBLAS with make flags DYNAMIC_ARCH=1 DYNAMIC_OLDER=1 USE_OPENMP=1 INTERFACE64=1 BINARY=64 NO_SHARED=0 NO_LAPACK=0 is it correct way to build openBLAS with int64 support ? Because i cannot see any 64 or _64 being appended or prefixed in library names or share object file names.

[access2rohit]

I need lapack interface so I added NO_LAPACK=0

[martin-frbg]

The suffix is a separate option (SYMBOLSUFFIX=64 , or LIBNAMESUFFIX=64 if you want only the library name but not the symbols to have a suffix)

[access2rohit]

So my build flags are correct ? But when I investigate the installation folder the lapacke_utils.h it shows int parameters in signatures and not lapack_int. Also, cblas.h shows blasint will INTERFACE64=1 BINARY=64 we sufficient to set that as int64_t as well ?

[access2rohit]

@martin-frbg i also saw the following code in file openblas_config.h

#ifdef OPENBLAS_USE64BITINT
typedef BLASLONG blasint;
#else
typedef int blasint;
#endif

I don't set OPENBLAS_USE64BITINT anywhere in my makefile or build commands

[martin-frbg]

The lapacke headers and their concept of "INTERFACE64" (there called LAPACK_ILP64) are copied verbatim from the "netlib"
reference implementation of LAPACK. While it is true that there are some int parameters left among the lapack_int, these are essentially flags like matrix_layout that are not expected to assume a wide range of values.
Regarding your other question, INTERFACE64=1 is converted to USE64BITINT internally (by a strangely complicated route - Makefile.system instructs the getarch tool to return that value), which gets OPENBLAS prepended in the installed version of the OpenBLAS headers.