Specialize pyarray_offset to be more amenable to constprop

[JamesWrigley]

This improves performance in tight loops, for me it pretty much fixes the performance hit compared to Array (#182 (comment)):

x_py = PyArray(np.random.rand(1000, 1000).T);
x = Array(x_py);

@benchmark sum($x)
BenchmarkTools.Trial: 10000 samples with 1 evaluation per sample.
 Range (min … max):  122.804 μs … 440.813 μs  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     124.024 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   125.858 μs ±   6.590 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

  ▂▆██▆▃▁        ▁     ▁▁▁▁▂▁▁▂▁▁                               ▂
  ██████████▇▇▆▆▇██▇▇█████████████████▇▇▇▆▆▇▆▆▆▆▄▅▅▆▅▄▄▅▄▅▄▅▂▅▅ █
  123 μs        Histogram: log(frequency) by time        146 μs <

 Memory estimate: 0 bytes, allocs estimate: 0.

@benchmark sum($x_py)
BenchmarkTools.Trial: 10000 samples with 1 evaluation per sample.
 Range (min … max):  128.664 μs … 447.404 μs  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     129.694 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   131.126 μs ±   4.813 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

   ▁▆██▆▂▁                            ▁▁▁▁▁▁▁▁▁                 ▂
  ▃████████▇▆▆▅▅▆▆▅▅▅▆▄▅▇█▇▅▃▄▄▆▆▇▇█████████████████▇▇█▇▇▇▇▆▆▆▆ █
  129 μs        Histogram: log(frequency) by time        143 μs <

 Memory estimate: 0 bytes, allocs estimate: 0.

[cjdoris]

Unfortunately this implementation is incorrect because PyArrays can be strided, even when linearly indexable.

Even more unfortunately, there is no room to encode the fact that the array might be contiguous in the type parameters, which means this cannot be fixed without a braking change.

Fortunately, the parameterisation has been changed in the v1 branch (

PythonCall.jl/src/API/types.jl

Line 53 in c0d44b8

"""

) and there is already a :contiguous flag implemented which does just this, so fingers crossed the performance is as you'd hope. Please give it a try (pkg> add PythonCall#v1) and let me know!

[JamesWrigley]

Ah, interesting. Indeed performance seems fixed on v1 :)