Is it possible to xt::adapt a buffer with strides?

[themightyoarfish]

I have an array of data elements of some size with padding, which I want to wrap with xt::adapt() such that the padding column is ignored. Context: each element is 3 floats for XYZ data, plus 1 float of padding for more efficient access.

I've tried a minimal program to get this to work, but it always terminates with

libc++abi: terminating with uncaught exception of type std::runtime_error: xbuffer_storage not resizable

Code:

#include <iomanip>
#include <vector>
#include <xtensor/xadapt.hpp>

int main(int argc, char *argv[]) {
  std::cout << std::fixed << std::setw(3) << std::setprecision(6)
            << std::setfill(' ');

  // make a vector of n_elems blocks with floats_per_elem * sizeof(float) bytes
  // each. Only the first n-1 of these are significant
  // Storage is row major, so elements are laid our one by one
  size_t n_elems = 5;
  size_t floats_per_elem = 4;
  std::vector<float> v(floats_per_elem * n_elems);
  std::cout << "Input array" << std::endl;
  for (int i = 0; i < v.size(); ++i) {
    // keep ever n-1th element at 0 and put some value in the others
    if (((i + 1) % floats_per_elem) != 0) {
      v[i] = 1 + 0.1 * i;
    }
    std::cout << v[i] << " ";
  }
  std::cout << std::endl;

  // desired xtensor shape: n_elems x floats_per_elem - 1, ignoring the padding
  // bytes
  std::vector<size_t> shape = {n_elems, floats_per_elem - 1};
  // rows are floats_per_elem elements apart
  std::vector<size_t> strides = {floats_per_elem, 1};

  auto adapted =
      xt::adapt(v.data(), v.size(), xt::no_ownership(), shape, strides);

  std::cout << "Adapted array" << std::endl;
  for (int i = 0; i < adapted.shape()[0]; ++i) {
    for (int j = 0; j < adapted.shape()[1]; ++j) {
      std::cout << adapted(i, j) << " ";
    }
    std::cout << std::endl;
  }
  return 0;
}

I build with

clang++ -o xtensor_test -I <deps>/include/ -std=c++20 xtensor_test.cpp  && ./xtensor_test

Do all xtensor arrays have to be contiguous, or is it possible to set up strides and shapes so that padding bytes are ignored?

[JohanMabille]

xt::adapt assumes the adapted container is contiguous. However, you can still achieve what you want, but with an extra step:

You can achieve what you want, but with an extra step:

std::vector<size_t> init_shape = {n_elems, floats_per_elem };
std::vector<size_t> shape = {n_elems, floats_per_elem - 1};
// rows are floats_per_elem elements apart
std::vector<size_t> strides = {floats_per_elem, 1};

auto adapted = xt::adapt(v.data(), v.size(), xt::no_ownership(), init_shape);
auto v = xt::strided_view(adapted, shape, strides);

// Or without intermediate adapted variable:
auto v = strided_view(xt::adapt(v.data(), v.size(), xt::no_ownership(), init_shape),
                      shape,
                      strides);

[github-actions]

This issue is stale because it has been open for 30 days with no activity.
It will be automatically closed in 14 days.