Ask any question about QuantStack projects: xtensor, xeus, voila, Jupyter widgets, mamba
Just trying the dirty hack
https://github.com/xtensor-stack/xtensor/blob/53ad31cf5d11b376de0d9e5eac91d6e7351c3cc3/include/xtensor/xsort.hpp#L155
does not properly resize the output
#define xtensor pytensor for fun gives a segmentation fault, becausehttps://github.com/xtensor-stack/xtensor/blob/53ad31cf5d11b376de0d9e5eac91d6e7351c3cc3/include/xtensor/xsort.hpp#L155
does not properly resize the output
we should run the xtensor test suite on bindings
So it seems that even though an intializer_list with one item is passed, https://github.com/xtensor-stack/xtensor/blob/53ad31cf5d11b376de0d9e5eac91d6e7351c3cc3/include/xtensor/xcontainer.hpp#L958
dim == 2??
Ah I see, the rank is not set to 1, i.e. this is not hit : https://github.com/xtensor-stack/xtensor/blob/53ad31cf5d11b376de0d9e5eac91d6e7351c3cc3/include/xtensor/xsort.hpp#L130
Calling with
#define xtensor pytensor and xt::xtensor<size_t, 2>
the underlying numpy array structure is a bit more involved than a C++ standard container, but I'm really not sure it would have a significative impact on performances
8 replies
Can someone look into xtensor-stack/xsimd#735, I'm willing to fix it if someone just points me to where.
It's reproducible on Godbolt, I also commented the link on the issue: https://godbolt.org/z/vj4fEGzqq
You can comment out the
#define then it works fine.
Another one... bool handling is wonky in general, this one doesn't need xsimd either. https://godbolt.org/z/bhh7Kfcen
Remove the true && to make it work.
when you write
true && (xt::view(a, 1) < 0.0), you create an unevaluated xfunction that cannot be cast to a bool
you need to evaluate it first:
true && (xt::view(a, 1) < 0.0)()
May I ask what the difference is between an
xfunction and an xexpression?
or said differently,
xexpression is a CRTP base class of xfunction
the issue was due to the way we adapt and handle strides
in xtensor, when a dimension's size is one, we set the stride to 0
it's not the case in numpy
I guess it should be pretty similar for xtensor (even simpler since it's a pure header library, nothing to build)
I think this is not yet implemented
but it shouldn't be too hard to do
do you want to take a stab at it? I'd be happy to help!
Is there a way to specify length when running load in xsimd? I have a data structure of 16 16-bit values I need to load. This would work fine in 128-bit and 256-bit architectures, but short of disallowing AVX512 at compile time, how do I make sure that doesnt result in an overflow and always uses 256-bits max?
Could anyone look at my implementation of pairwise distance (just two lines) and tell me why it doesn't execute as fast as a simple for-loop?
auto diff = xt::view(in1, xt::all(), xt::newaxis(), xt::all())
- xt::view(in2, xt::newaxis(), xt::all(), xt::all());
xt::view(state.pairwise_distance, xt::all(), xt::all())
= xt::sqrt(xt::sum(diff * diff, { 2 }));