Commit Graph

76 Commits

Author SHA1 Message Date
dzaima
52fb05fcd4 unify ⌾(a⊸⊏) typed array paths, fix csz==0 2024-05-20 01:03:57 +03:00
Andrea Piseri
60a28a40ca Clean up collision checking, use vector copy functions 2024-05-19 16:59:55 +02:00
dzaima
9f6151d92d expand ⌾(a⊸⊏) tests, stylistic changes/cleanup 2024-05-19 02:14:08 +03:00
Andrea Piseri
99a23a97d8 Fix typo in the error message 2024-05-19 00:10:28 +02:00
Andrea Piseri
3738d0d65d Formatting 2024-05-18 23:58:11 +02:00
Andrea Piseri
93188a9f90 Use COPY_TO and EQFN in the typed array code path 2024-05-18 23:56:13 +02:00
Andrea Piseri
06808414da Separate code path for cxsz=1, fix out of bounds read for RNK(x)==0 2024-05-18 21:45:18 +02:00
Andrea Piseri
7f28308e44 native path in select_ucw for high rank 𝕩
`𝔽⌾(a⊸⊏)𝕩` now does not need to go through the self-hosted runtime if
`1<=𝕩`. Instead the `select_replace` helper function is parametrized
over the length of `𝕩` (`xl`) and item amount of the cell of `𝕩` (`xcia`).

- The `EQ` macro is modified to not immediately mark the cell as
  populated, so that multiple replacements can be done on the cell
  on the first assignment to it.

- The `DONE_CW` macro is invoked to mark the current cell as populated
  when every element in it has been assigned.

- A loop over the cell contents is introduced to copy the elements in
  `𝕩`. This should be fine as it is an easily predictable jump,
  but a performance regression is possible and a separate code path
  could be introduced in the future.

The change introduces more extensive checking on the shape of `𝔽`'s
result, as for high rank `𝕩` the requirement should be `(≢𝔽a⊏𝕩)≡(≢a)∾1↓≢𝕩`.

The old behaviour of `select_replace` is recovered by passing `xl=xia`
and `xcia=1` in the implementation of `pick_ucw`.
2024-05-18 12:12:17 +02:00
dzaima
ac069134c3 fix ⟨1.5⟩ ⊏ "ab"‿"cd" 2024-03-25 16:18:01 +02:00
dzaima
a94903cb12 fix 𝕨⊏unit 2024-03-25 16:18:01 +02:00
dzaima
c078f62b1f add GLOBAL, INIT_GLOBAL, or const to globals 2024-02-01 21:13:39 +02:00
dzaima
f0c1962596 more typed ⌾(list⊸⊏) 2023-12-19 17:41:17 +02:00
dzaima
8c7a3be036 remove unused includes 2023-11-13 21:48:01 +02:00
dzaima
5efd90cc91 rename elWidthLogBits to something less long & add byte variant 2023-09-12 19:58:56 +03:00
dzaima
78bef5fe32 consistent 𝕨⊏𝕩 error messages across argument types 2023-05-30 14:28:36 +03:00
dzaima
df348896fd add a NOGC_E for conflicting⌾(list⊸⊏) generic 2023-05-19 22:51:02 +03:00
dzaima
c1488cbf72 faster arr⌾(is⊸⊏)list 2023-05-19 14:15:25 +03:00
dzaima
0572fcc5b2 native F⌾(list⊸⊑) & ⌾(listOfLists⊸⊑) 2023-05-18 01:54:01 +03:00
dzaima
0598ce8357 use getFillQ less 2023-05-09 00:14:14 +03:00
dzaima
bb7c37960e a round of more incG/incByG 2023-04-29 17:39:13 +03:00
dzaima
3b580cbd0c rename unit creation methods 2023-04-29 17:39:13 +03:00
dzaima
865257fdb6 more header file cleanup 2023-04-29 17:39:13 +03:00
dzaima
1c8916949e fix ⊏ with odd shape 𝕩 & depth 2 𝕨 2023-04-14 00:14:57 +03:00
dzaima
9951f20751 clean up some error messages 2023-04-08 22:12:15 +03:00
dzaima
3a45eac1aa separate SSE2, implement SSE2 dyarith & bitsel
also separate maskstore to homogeneous & top bit mask variants, and make blend require equal element widths
2023-04-01 20:40:23 +03:00
dzaima
1f70c08aa2 fillarr NOGC 2023-03-26 15:43:59 +03:00
dzaima
8194f51b30 MAKE_MUT_INIT, move mut_init to mut.c 2023-03-11 16:52:09 +02:00
dzaima
01c457b77b make all cpy.*Arr function results Arr* 2023-03-11 16:51:13 +02:00
dzaima
deee2c7a18 use C1 & C2 more 2023-02-24 15:53:11 +02:00
dzaima
855541954f initial bit widen&narrow for ⊏ 2023-02-10 17:59:22 +02:00
dzaima
23d99d7a24 more precise detection of when bit 𝕩 won't use simd for widening path 2023-02-04 18:17:21 +02:00
dzaima
f39ca95c58 maybe slightly better bit→i8→bit heuristic
the more proper thing is to pad cells to the nearest power-of-two byte size anyway
2023-02-04 17:16:01 +02:00
dzaima
be9d0c287e better 𝕨⊏𝕩 with non-power-of-two cell sizes 2023-02-04 03:30:05 +02:00
dzaima
d2679da7e8 ⊏: don't get in an infinite loop of f64 𝕨 2023-02-04 03:04:06 +02:00
dzaima
98f6944440 unify rank 1 and high-rank ⊏ 2023-02-03 18:55:19 +02:00
dzaima
05c1270344 better 𝕨⊏𝕩 with 0=×´≢𝕨 2023-02-02 00:40:52 +02:00
dzaima
b1cace31ac don't make a slice on n⊏list 2023-01-22 19:31:11 +02:00
dzaima
013897cc16 add runtime calls to WARN_SLOW 2022-12-26 23:15:56 +02:00
dzaima
bb7fe02a45 separate checks for x86-64 & NEON Singeli importing
& restrict the processed Singeli files on NEON to the ones supported
2022-12-24 15:27:53 +02:00
dzaima
09b3279958 unified Singeli including mechanism 2022-12-18 01:54:07 +02:00
dzaima
7c6676a492 copy over custom _mm_loadu_si32 for old gcc 2022-12-01 13:37:01 +02:00
Marshall Lochbaum
4077596b89 Sparse initialization for ⌾(i⊸⊏) byte table 2022-11-30 16:32:06 -05:00
Marshall Lochbaum
d7a06befb5 Select implementation comments 2022-11-30 16:32:06 -05:00
Marshall Lochbaum
534c92f38a Select from <=128 booleans with shuffles 2022-11-29 22:14:57 -05:00
Marshall Lochbaum
c17448ed71 Faster bit selection 2022-11-29 11:29:25 -05:00
Marshall Lochbaum
7cb8cf5772 Enable selection with AVX2 gather for i8 to i32 indices and 1- to 8-byte values 2022-11-27 09:18:21 -05:00
dzaima
8d6172126c use void* instead of u8* for singeli function arguments 2022-10-25 17:27:08 +03:00
dzaima
66584ce149 fix 0‿¯1⊏129⥊100 2022-10-21 18:17:12 +03:00
dzaima
59cadb06fe fix ⟨⟨⟩⟩⊏⟨⟩ incorrectly erroring
incl. ⟨⟨⟩,⟨⟩⟩⊏0‿0⥊⟨⟩, ⟨⟨⟩⟩⊏0‿0⥊⟨⟩, etc
2022-09-08 00:56:31 +03:00
dzaima
66d32e0acb fast exit for depth-2 𝕨 in 𝕨⊏𝕩 2022-09-08 00:47:25 +03:00