On Sunday, 6 December 2020 at 17:30:13 UTC, data pulverizer wrote:
On Saturday, 5 December 2020 at 07:44:33 UTC, 9il wrote:
sweep_ndslice uses (2*N - 1) arrays to index U, this allows
LDC to unroll the loop.
For example, for 2D case, withNeighboursSum [2] will store the
pointer to the result, and the pointer at rows above and below.
matrix:
--------------
------a------- above iterator
------r------- the result
------b------- below iterator
--------------
Also, for AVX-512 targets it allows vectorizing the loop [1].
The benchmark has been run on the AVX2 CPU.
[1] https://github.com/typohnebild/numpy-vs-mir/issues/4
[2]
http://mir-algorithm.libmir.org/mir_ndslice_topology.html#.withNeighboursSum
Very interesting, thank you for the explanations. Are there
journal/book other implementation references for these
approaches to implementing tensor-like multidimensional arrays?
I don't know. Tensors aren't so complex. The complex part is a
design that allows Mir to construct and iterate various kinds of
lazy tensors of any complexity and have quite a universal API,
and all of these are boosted by the fact that the user-provided
kernel(lambda) function is optimized by the compiler without the
overhead.
