Speed up Image.getbbox(); fix it for I;16 modes

[akx]

Hoist/restrict and autovectorization optimizations for getbbox().

A review also showed that I;16 modes weren't correctly supported by this function (they were iterating over int8), so that got fixed too. I know it's bad form to mix optimizations and bugfixes into one PR, but the bugfix would have required the const type* row = ... hoist assignments anyhow.

-------------------------- benchmark 'bbox': 20 tests, 2 sources ---------------------------
Name (time in us)                               0001_40ea9a1 Min  0002_d07694f Min      ΔMin
--------------------------------------------------------------------------------------------
test_getbbox[1024x1024-L-half-corner]                   226.7091           10.9579    -95.2%
test_getbbox[1024x1024-RGB-half-corner]                 229.2499           38.2080    -83.3%
test_getbbox[1024x1024-RGBA-half-corner]                229.3329           40.5840    -82.3%
test_getbbox[1024x1024-LA-half-corner]                  230.6249           38.2080    -83.4%
test_getbbox[1024x1024-I;16-half-centered]              231.0410           21.3749    -90.7%
test_getbbox[1024x1024-L-half-centered]                 240.4579           11.3329    -95.3%
test_getbbox[1024x1024-RGBA-half-centered]              249.7090           44.7499    -82.1%
test_getbbox[1024x1024-LA-half-centered]                249.8330           41.2500    -83.5%
test_getbbox[1024x1024-RGB-half-centered]               251.2090           41.9160    -83.3%
test_getbbox[1024x1024-L-empty]                         295.4580           11.4590    -96.1%
test_getbbox[1024x1024-L-small-centered]                297.5831           12.5830    -95.8%
test_getbbox[1024x1024-LA-empty]                        298.1250           43.9170    -85.3%
test_getbbox[1024x1024-RGB-empty]                       298.1250           47.5419    -84.1%
test_getbbox[1024x1024-RGBA-empty]                      298.1670           47.0839    -84.2%
test_getbbox[1024x1024-I;16-small-centered]             299.1251           24.1250    -91.9%
test_getbbox[1024x1024-I;16-empty]                      300.4170           21.9160    -92.7%
test_getbbox[1024x1024-I;16-half-corner]                301.2920           19.0841    -93.7%
test_getbbox[1024x1024-LA-small-centered]               304.9170           51.2920    -83.2%
test_getbbox[1024x1024-RGB-small-centered]              305.0419           54.7920    -82.0%
test_getbbox[1024x1024-RGBA-small-centered]             305.0419           50.4589    -83.5%
--------------------------------------------------------------------------------------------

[codspeed]

Merging this PR will improve performance by 69.8%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 20 improved benchmarks
✅ 583 untouched benchmarks
⏩ 335 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_getbbox[1237x811-L-empty]	3 ms	1.1 ms	×2.9
⚡	test_getbbox[1237x811-L-small-centered]	3 ms	1.1 ms	×2.9
⚡	test_getbbox[1237x811-L-half-centered]	2,353.5 µs	857.1 µs	×2.7
⚡	test_getbbox[1237x811-L-half-corner]	2,331.3 µs	852.9 µs	×2.7
⚡	test_getbbox[1237x811-I;16-half-corner]	3.1 ms	1.6 ms	+95.03%
⚡	test_getbbox[1237x811-I;16-empty]	3.1 ms	2 ms	+51.17%
⚡	test_getbbox[1237x811-I;16-small-centered]	3.1 ms	2 ms	+50.84%
⚡	test_getbbox[1237x811-I;16-half-centered]	2.3 ms	1.6 ms	+48.4%
⚡	test_getbbox[1237x811-RGB-empty]	6 ms	4.1 ms	+46.94%
⚡	test_getbbox[1237x811-LA-empty]	6 ms	4.1 ms	+46.93%
⚡	test_getbbox[1237x811-RGBA-empty]	6 ms	4.1 ms	+46.93%
⚡	test_getbbox[1237x811-RGBA-small-centered]	5.9 ms	4 ms	+46.9%
⚡	test_getbbox[1237x811-RGB-small-centered]	5.9 ms	4 ms	+46.9%
⚡	test_getbbox[1237x811-LA-small-centered]	5.9 ms	4 ms	+46.85%
⚡	test_getbbox[1237x811-RGBA-half-centered]	4.5 ms	3.1 ms	+46.47%
⚡	test_getbbox[1237x811-RGB-half-centered]	4.5 ms	3.1 ms	+46.47%
⚡	test_getbbox[1237x811-LA-half-centered]	4.5 ms	3.1 ms	+46.46%
⚡	test_getbbox[1237x811-RGB-half-corner]	4.5 ms	3.1 ms	+45.77%
⚡	test_getbbox[1237x811-RGBA-half-corner]	4.5 ms	3.1 ms	+45.77%
⚡	test_getbbox[1237x811-LA-half-corner]	4.5 ms	3.1 ms	+45.76%

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing akx:unbound (98cfb1b) with main (3bcb709)}

Footnotes

1. 335 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[radarhere]

@akx I would appreciate it if you didn't rebase after an approval. I no longer find it easy to know what hugovk approved, and what he didn't

[akx]

@radarhere Sorry, I'll try to keep that in mind! 👍

git range-diff for the two commits in the above "@akx force-pushed from 9f43a82 to 98cfb1b" message show the patches are equivalent, so there's nothing that Hugo didn't approve (or, in the case of 98cfb1b, didn't suggest):

$ git range-diff --left-only 9f43a8289...98cfb1b48
  1:  e782d06e0 = 179:  122663023 Add additional tests for bbox
  2:  3ef06b2ee = 180:  a220ea4b3 Apply hoist optimizations to getbbox
  3:  efe739148 = 181:  343ad3ce1 Fix getbbox for I;16 modes
  4:  4ec1a3d46 = 182:  b1f18fefd Speed up `ImagingGetBBox` by doing vectorisable work upfront
  5:  2a5ef4c31 = 183:  41a3c7fd1 Reword comment
  6:  229e2932b = 184:  e78eac080 Test I;16N too
  7:  9f43a8289 = 185:  98cfb1b48 Check image type, not mode specifically