verify NEON.update failure in recent commit(at least from v3.11.0) about embree-aarch64

[syoyo]

The issue only happens when EMBREE_RAY_PACKETS=On

[maikschulze]

I can confirm that NEON.update used to work and now fails.

For no particular reason, I've selected to track the introduction of FAILED for
std::vector<std::string> arguments = { "", "--no-colors", "--run .*SSE2.update.dynamic.DynamicFast.LowQuality.Intersect4.*", "--sequential" };
, which is 1 of 120 update tests now failing.

I can confirm that:
8246a54 PASSED
4ecf706 PASSED
9133976 PASSED
d6185d1 PASSED
5f3350f FAILED
ba6d997 FAILED

I briefly debugged into the failure but could not pinpoint the cause. There are minute numerical differences compared to SSE2 that eventually lead to different branching, but nothing particular stands point.

Any idea what might cause it?

[maikschulze]

I was able to pin-point the failure to a single line of code that was introduced with 5f3350f

I ran the 'verify' suite before with 316c2fb (left) and after (right) the reversal with packets enabled:

verify_before_sse2_with_packets.txt
verify_before_avx2_with_packets.crashlog.txt

verify_after_sse2__with_packets.txt
verify_after_avx2__with_packets.crashlog.txt

For SSE2 / NEON, all tests will pass.

Since the reversal will impact performance to a significant degree, I refrain from making such a merge request right away. I suggest to investigate the numerical behavior over the next days. Maybe we come up with a better solution.

I believe the "AVX2 emulation"-related crashes are caused by access to functions that are not present because of the absence of the "AVX" code path. We may have to improve the function dispatch for NEON2X in this regard. I will debug into this more closely.

[Developer-Ecosystem-Engineering]

Patch incoming

Fixed in #41

[syoyo]

PR #44 has been merged and update test in verify all passes for every ISA configuration(NEON, AVX). So I think its ready to close this issue.

[maikschulze]

I agree, I also think this issue has been fixed. Many thanks to @Developer-Ecosystem-Engineering !

[syoyo]

@maikschulze @Developer-Ecosystem-Engineering 🎉

I have updated README: 54a6ebc