Building with WITH_CYCLES_NATIVE_ONLY on AVX2+MSVC system leads to a bad combination of sse2 kernel+BVH8.
What is happening is:
- MSVC does not have an -march=native switch, so kernel.cpp gets build with the default flags and you and up with a run of the mill SSE2 kernel.
- DebugFlags().cpu.has_avx2() is true, and it decides BVH8 is a valid option, normally not an issue, since kernel.cpp detects the AVX2 as well and has the right kernel.
- But not for msvc which tries to match up and sse2 kernel with BVH8 and thing go boom