bsplines

working title

bsplines's People

Contributors

out of range behavior options in 1-d spline: constant (default= knot value), extrapolate, error
error messaging
prettify exterior behavior and boundary conditions parameters in Python interface
more C tests for basis functions
continuous 3rd derivative ("not-a-knot") BC option
uniform 1-d spline
2-d spline
handle constant extend (different from value for 2-d)
vector derivative BC in 2-d spline
need to use np.require for C continguousness? No.
checks for length of input arrays
- minimum: 2 points
- one not-a-knot: need 3 points
- both not-a-knot: need 4 points (with 3 points, left and right conditions are the same).
add Python tests from scipy interpolate
benchmark 2-d spline
uniform 2-d spline
SIMD
- check AVX at runtime: stack overflow crc32 example
- check at compile time: $ gcc -march=native -dM -E - < /dev/null | egrep "SSE|AVX" | sort shows that __AVX__ will be defined by gcc.
- AVX matrix multiply example
- AVX intrinsics: https://software.intel.com/en-us/node/524038
- _mm256_loadu_pd load 4 doubles
- _mm256_mul_pd vector multiply
- _mm256_hadd_pd add adjacent pairs
- SO answer on _mm256 dot product
- checking AVX at runtime might not work because if it doesn't exist, compiling with AVX support might generate code elsewhere (besides intrinsics) that depends on AVX.