GEMM/IGEMM tests not working as expected for RVV?

[ken-unger]

@dsharlet in PR#7932 in response to my noting that the unit test didn't catch my dumb rvv igemm bug, you said;

Yes it does seem weird the unit tests would not catch this. I am pretty sure the unit tests catch far more subtle bugs than this. I suspect there's a problem specific to RVV, could you please take a look? There's a lot that is unique to the RVV tests due to the non-constant SIMD length.

I started looking into this a bit, and for sure there seems to be some problem. I've noted that qd8-f32-qc8w-igemm-minmax-test and qu8-igemm-minmax-fp32-test (an upcoming RVV PR) do not actually execute any of the tests claimed to pass! In other gemm & igemm tests I see very few RVV tests actually executed, and in some cases, notably qu8-igemm-minmax-fp32-test, even the scalar tests do not appear to be actually executed.

So far just simplistic debugging, with a couple of printfs to see if it ever makes it here https://github.com/google/XNNPACK/blob/master/test/gemm-microkernel-tester.cc#L80

I'll look at this further in the next few days but thought I'd log this issue to track,

[ken-unger]

I'm probably not understanding something correctly, since this has been here for a while but shouldn't https://github.com/google/XNNPACK/blob/master/test/gemm-microkernel-tester.cc#L70

for (size_t bl = params.loop_bl_.from; bl <= tester.k() / 2; bl = params.loop_bl_.next(bl)) {

be this instead? Or at least this solves the issue.

for (size_t bl = params.loop_bl_.from; bl <= (tester.k() + 1) / 2; bl = params.loop_bl_.next(bl)) {

[CNOCycle]

This issue has been reported in #7949. Some propoposed solutions were proposed and await further discussion.