[Numpy-svn] [numpy/numpy] 46dd68: BUG/ENH: blasdot: make *dot calls 64-bit safe
GitHub
noreply at github.com
Mon Dec 30 05:23:49 EST 2013
Branch: refs/heads/master
Home: https://github.com/numpy/numpy
Commit: 46dd681ff4aed02a87b27db684f8600b5b7a7fa3
https://github.com/numpy/numpy/commit/46dd681ff4aed02a87b27db684f8600b5b7a7fa3
Author: Lars Buitinck <larsmans at gmail.com>
Date: 2013-11-21 (Thu, 21 Nov 2013)
Changed paths:
M numpy/core/blasdot/_dotblas.c
Log Message:
-----------
BUG/ENH: blasdot: make *dot calls 64-bit safe
BLAS uses int for all its size types, so the size is going to be
silently truncated on platforms where sizeof npy_intp > sizeof int,
i.e. on modern 64-bit platforms. In this case, we do a dot product
in reasonably-sized chunks and add the result ourselves.
On platforms where sizeof(int) == sizeof(npy_intp), i.e. typical
32-bit platforms, it's up to the compiler to optimize the loop away.
Also tried to make the code more readable, replaced int by npy_intp
in a few places and removed useless register declarations (modern
compilers ignore those).
Commit: 1d1803257467c554db208a24b857d613b872b9bd
https://github.com/numpy/numpy/commit/1d1803257467c554db208a24b857d613b872b9bd
Author: Lars Buitinck <larsmans at gmail.com>
Date: 2013-11-21 (Thu, 21 Nov 2013)
Changed paths:
M numpy/core/blasdot/_dotblas.c
Log Message:
-----------
ENH: dotblas: re-route all cblas_?dot calls to chunked versions
Also factored out and simplified NumPy->BLAS stride conversion.
Commit: 14fb7000694523919fe4045096fbfa3703b1d1fe
https://github.com/numpy/numpy/commit/14fb7000694523919fe4045096fbfa3703b1d1fe
Author: seberg <sebastian at sipsolutions.net>
Date: 2013-12-30 (Mon, 30 Dec 2013)
Changed paths:
M numpy/core/blasdot/_dotblas.c
Log Message:
-----------
Merge pull request #3916 from larsmans/dotblas-64bit-safe
64-bit safe _dotblas
Compare: https://github.com/numpy/numpy/compare/61998c22df08...14fb70006945
More information about the Numpy-svn
mailing list