Regenerate binaries on ISPC 1.29.1

[MarijnS95]

https://github.com/ispc/ispc/releases/tag/v1.29.1

TODO: Still need to compare performance, but perhaps this helps on newer architectures. Might also have to evaluate if we're simply missing some TargetISA flags relevant for newer SoCs?

[MarijnS95]

Turns out there are a bunch of new generic target ISAs to streamline which vector sizes/widths to select, as well as Apple-specific CPU targets :)

[MarijnS95]

On the MacBook Air M4

Main @ 6e7b616 (ISPC 1.20...)

Downsample `square_test.png` using ispc_downsampler
                        time:   [38.827 ms 38.848 ms 38.884 ms]
Found 5 outliers among 100 measurements (5.00%)
  3 (3.00%) high mild
  2 (2.00%) high severe

This ispc-1.27 PR @ 3556673

Downsample `square_test.png` using ispc_downsampler
                        time:   [48.220 ms 48.253 ms 48.287 ms]
                        change: [+24.103% +24.190% +24.278%] (p = 0.00 < 0.05)
                        Performance has regressed.

Recompiling locally on the M4 Air (ispc 1.27.0 from brew using cargo b -rF ispc):

Downsample `square_test.png` using ispc_downsampler
                        time:   [46.576 ms 46.586 ms 46.596 ms]
                        change: [-3.5237% -3.4550% -3.3855%] (p = 0.00 < 0.05)
                        Performance has improved.

That's a significant performance deficit, which we should investigate before merging. Even playing around with the new CPU flags from Twinklebear/ispc-rs#42, or the generic ISAs, or removing .target_isas() altogether to compile natively for the host yields no improvement.

Funny thing is, with the ISPC test this M4 Air whines a little, but it doesn't during resize 😓

[MarijnS95]

Looks like performance is not restored in 1.28, or we're still doing something wrong. Barely any change compared against 1.27 (which was 24% slower than main per the above):

This PR @ 00d0256

❯ cargo bench
Downsample `square_test.png` using ispc_downsampler
                        time:   [48.412 ms 48.475 ms 48.539 ms]
                        change: [+29.098% +29.471% +29.821%] (p = 0.00 < 0.05)
                        Performance has regressed.

[MarijnS95]

Re-running this test on my host, recompiled on this ISPC version:

❯ ispc --version
Intel(r) Implicit SPMD Program Compiler (Intel(r) ISPC), 1.28.2 (build commit  @ 20250924, LLVM 20.1.8)

On latest main @ f2ddfab (but not using those prebuilts)

❯ cargo bench
Downsample `square_test.png` using ispc_downsampler
                        time:   [46.776 ms 46.875 ms 46.969 ms]

Then following the suggestion from @Jasper-Bekkers in Traverse-Research/intel-tex-rs-2#42 to only use i32x4 because NEON is 128-bits slightly regresses performance:

❯ cargo bench
Downsample `square_test.png` using ispc_downsampler
                        time:   [48.003 ms 48.101 ms 48.196 ms]
                        change: [+2.3395% +2.6161% +2.9034%] (p = 0.00 < 0.05)
                        Performance has regressed.

Also, this M4 chip is supposed to save SME (Scalable Matrix Extensions) but not SVE (Scalable Vector Extensions) and confirmed with sysctl -a hw.optional (and NEON is confirmed as well).

Perhaps this needs to be reported upstream as I'm slightly out of ideas how to best bisect this compiler performance regression.

[MarijnS95]

Just went back in history to generate the blobs for all missing versions:

ISPC 1.23 @ 754d4bf

Downsample `square_test.png` using ispc_downsampler
                        time:   [37.430 ms 37.550 ms 37.666 ms]

ISPC 1.24 @

Downsample `square_test.png` using ispc_downsampler
                        time:   [37.180 ms 37.317 ms 37.454 ms]
                        change: [-1.1514% -0.6188% -0.1473%] (p = 0.01 < 0.05)
                        Change within noise threshold.

ISPC 1.25.3

Downsample `square_test.png` using ispc_downsampler
                        time:   [38.024 ms 38.151 ms 38.315 ms]
                        change: [+1.6876% +2.2352% +2.8037%] (p = 0.00 < 0.05)
                        Performance has regressed.

ISPC 1.26

Downsample `square_test.png` using ispc_downsampler
                        time:   [49.251 ms 49.422 ms 49.588 ms]
                        change: [+29.523% +30.093% +30.690%] (p = 0.00 < 0.05)
                        Performance has regressed.

1.26 is where this regression happened.

Turns out that 1.26 release is exactly where a bunch of Apple improvements have been announced. Unfortunately, playing with that new --darwin-version-min flag, or the new CPU targets (which are only available up to A17, the "predecessor" to M4 in the iPhone space) mentioned above, don't make a difference. I couldn't immediately find if those iPhone skews have support for vector extensions at all..?

[Jasper-Bekkers]

Yeah I closed thar PR because later I realized why there was a big delta: I was profiling on battery.

[MarijnS95]

@Jasper-Bekkers Oh I'm also exclusively developing on battery (the perks of Apple putting RTGs in these MacBooks 🤤) but the ±37ms vs ±45ms regression remains consistent.