@alalek, we have v_load_aligned() to load from aligned addresses. v_load() must be able to handle unaligned data.

@vpisarev This is not about SIMD only - check #16139 and related issue.
Also there is no way to propagate alignment requirements through template.

v_load_aligned() to load from aligned addresses

... aligned on all 128 / 256 bits. This requirement is much stronger than alignment on lane(base) type.

@alalek, oh, I see, I misunderstood the code. But can we replace CV_Assert() with CV_DbgAssert()? Since adding the check to each load and store operation is quite significant overhead.

replace CV_Assert() with CV_DbgAssert()

It is intrin_cpp.hpp file. This SIMD implementation is not used by default:

• tests for "intrin_cpp" are enabled 2.5 months ago: core(test): extending tests with SIMD C++ emulation code (intrin_cpp.hpp) #15925
• for other OpenCV code it is enabled 1 month ago: core: fix simd emulator code #16236 (through C++ flag)

There are weekly "SIMD emulator" builders which run intrin_cpp code through all tests (to ensure that SIMD functions are used properly).
Replacing to CV_DbgAssert() would require to switch these builders in debug mode - this is much slower than current checks (+emulated SIMD code) in release mode. So I would like to avoid that.

👍