ENH: Implement essential intrinsics required by the upcoming SIMD optimizations(0) #22306
Conversation
max/min that reverse IEC 60559's NaN beahvior(propagates NaNs) for float data types
npyv_maxn_##SFX
npyv_minn_##SFX
max/min reduction for all float and integer vector data types
npyv_reduce_max_##SFX
npyv_reduce_min_##SFX
max/min reduction supports IEC 60559 for float data types
npyv_reduce_maxp_##SFX
npyv_reduce_minp_##SFX
max/min reduction reverse IEC 60559's NaN beahvior(propagates NaNs) for float data types
npyv_reduce_maxn_##SFX
npyv_reduce_minn_##SFX
also, this patch implements new intrinsics to extract the first vector lane:
npyv_extract0_##SFX
npyv_extract0_##SFX
57e0148 to
e3ad145
Compare
npyv_any_##SFX: returns true if any of the elements is not equal to zero npyv_all_##SFX: returns true if all elements are not equal to zero
e3ad145 to
6ef4c8b
Compare
|
cc @mattip |
|
The errors are unrelated, they are on account of log length limitations on travis. I made a conservative fix for that before, looks like it needs to be more drastic. |
|
close/reopen |
| continue | ||
| _min = self.min(vdata_a, vdata_b) | ||
| assert _min == data_min | ||
| chk_nan = {"xp": 1, "np": 1, "nn": 2, "xn": 2}.get(intrin[-2:], 0) |
There was a problem hiding this comment.
Would probably be clearer as part of the parametrize but doesn't matter. I am mostly curious what min/max implement for float values? Does the result depend on the order?
In any case, looked at the code and does look good to me (not that I am very fluid at simd). Not sure if the tests cover all the permutations they could for reductions, but I also trust our integration tests for that.
@mattip will you have another quick look?
There was a problem hiding this comment.
I am mostly curious what min/max implement for float values? Does the result depend on the order?
no, it doesn't, just check the tail of the intrinsic name to determine the NaN behavior.
Not sure if the tests cover all the permutations they could for reductions
I'm positive about it.
There was a problem hiding this comment.
I'm working on refactoring the whole testing unit starting from _simd module to count more on parametrizing rather than inheritance.
see the new numpy/core/tests/test_simd.py part of #21057
|
Thanks @seiko2plus |
This pullrequest adds:
intrinsics to check true cross all vector lanes
npyv_any_##SFX: returns true if any of the elements is not equal to zero
npyv_all_##SFX: returns true if all elements are not equal to zero
max/min that reverse IEC 60559's NaN behavior(propagates NaNs) for float data types
npyv_maxn_##SFX
npyv_minn_##SFX
max/min reduction for all float and integer vector data types
npyv_reduce_max_##SFX
npyv_reduce_min_##SFX
max/min reduction supports IEC 60559 for float data types
npyv_reduce_maxp_##SFX
npyv_reduce_minp_##SFX
max/min reduction reverse IEC 60559's NaN behavior(propagates NaNs) for float data types
npyv_reduce_maxn_##SFX
npyv_reduce_minn_##SFX
intrinsics to extract the first vector lane:
npyv_extract0_##SFX
npyv_extract0_##SFX
And removes: